├── .gitignore ├── LICENSE ├── Supplymental_material.pdf ├── dataset ├── DND_sRGB.py ├── SIDD_sRGB.py ├── __init__.py └── base.py ├── image_denoising.yaml ├── model ├── __init__.py ├── base.py ├── bnn.py └── three_stage.py ├── network ├── __init__.py ├── bnn.py ├── lan.py └── unet.py ├── option ├── bnn.json └── three_stage.json ├── pretrained_models ├── BNN.pth ├── LAN.pth └── UNet.pth ├── readme.md ├── submit_test ├── ensemble_wrapper.py ├── test_DND.py └── test_SIDD.py ├── utils ├── build.py ├── io.py └── option.py └── validate └── validate_SIDD.py /.gitignore: -------------------------------------------------------------------------------- 1 | .xml 2 | .idea 3 | .idea/workspace.xml 4 | .DS_Store 5 | */__pycache__git 6 | .pyc 7 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | GNU GENERAL PUBLIC LICENSE 2 | Version 3, 29 June 2007 3 | 4 | Copyright (C) 2007 Free Software Foundation, Inc. 5 | Everyone is permitted to copy and distribute verbatim copies 6 | of this license document, but changing it is not allowed. 7 | 8 | Preamble 9 | 10 | The GNU General Public License is a free, copyleft license for 11 | software and other kinds of works. 12 | 13 | The licenses for most software and other practical works are designed 14 | to take away your freedom to share and change the works. By contrast, 15 | the GNU General Public License is intended to guarantee your freedom to 16 | share and change all versions of a program--to make sure it remains free 17 | software for all its users. We, the Free Software Foundation, use the 18 | GNU General Public License for most of our software; it applies also to 19 | any other work released this way by its authors. You can apply it to 20 | your programs, too. 21 | 22 | When we speak of free software, we are referring to freedom, not 23 | price. Our General Public Licenses are designed to make sure that you 24 | have the freedom to distribute copies of free software (and charge for 25 | them if you wish), that you receive source code or can get it if you 26 | want it, that you can change the software or use pieces of it in new 27 | free programs, and that you know you can do these things. 28 | 29 | To protect your rights, we need to prevent others from denying you 30 | these rights or asking you to surrender the rights. Therefore, you have 31 | certain responsibilities if you distribute copies of the software, or if 32 | you modify it: responsibilities to respect the freedom of others. 33 | 34 | For example, if you distribute copies of such a program, whether 35 | gratis or for a fee, you must pass on to the recipients the same 36 | freedoms that you received. You must make sure that they, too, receive 37 | or can get the source code. And you must show them these terms so they 38 | know their rights. 39 | 40 | Developers that use the GNU GPL protect your rights with two steps: 41 | (1) assert copyright on the software, and (2) offer you this License 42 | giving you legal permission to copy, distribute and/or modify it. 43 | 44 | For the developers' and authors' protection, the GPL clearly explains 45 | that there is no warranty for this free software. For both users' and 46 | authors' sake, the GPL requires that modified versions be marked as 47 | changed, so that their problems will not be attributed erroneously to 48 | authors of previous versions. 49 | 50 | Some devices are designed to deny users access to install or run 51 | modified versions of the software inside them, although the manufacturer 52 | can do so. This is fundamentally incompatible with the aim of 53 | protecting users' freedom to change the software. The systematic 54 | pattern of such abuse occurs in the area of products for individuals to 55 | use, which is precisely where it is most unacceptable. Therefore, we 56 | have designed this version of the GPL to prohibit the practice for those 57 | products. If such problems arise substantially in other domains, we 58 | stand ready to extend this provision to those domains in future versions 59 | of the GPL, as needed to protect the freedom of users. 60 | 61 | Finally, every program is threatened constantly by software patents. 62 | States should not allow patents to restrict development and use of 63 | software on general-purpose computers, but in those that do, we wish to 64 | avoid the special danger that patents applied to a free program could 65 | make it effectively proprietary. To prevent this, the GPL assures that 66 | patents cannot be used to render the program non-free. 67 | 68 | The precise terms and conditions for copying, distribution and 69 | modification follow. 70 | 71 | TERMS AND CONDITIONS 72 | 73 | 0. Definitions. 74 | 75 | "This License" refers to version 3 of the GNU General Public License. 76 | 77 | "Copyright" also means copyright-like laws that apply to other kinds of 78 | works, such as semiconductor masks. 79 | 80 | "The Program" refers to any copyrightable work licensed under this 81 | License. Each licensee is addressed as "you". "Licensees" and 82 | "recipients" may be individuals or organizations. 83 | 84 | To "modify" a work means to copy from or adapt all or part of the work 85 | in a fashion requiring copyright permission, other than the making of an 86 | exact copy. The resulting work is called a "modified version" of the 87 | earlier work or a work "based on" the earlier work. 88 | 89 | A "covered work" means either the unmodified Program or a work based 90 | on the Program. 91 | 92 | To "propagate" a work means to do anything with it that, without 93 | permission, would make you directly or secondarily liable for 94 | infringement under applicable copyright law, except executing it on a 95 | computer or modifying a private copy. Propagation includes copying, 96 | distribution (with or without modification), making available to the 97 | public, and in some countries other activities as well. 98 | 99 | To "convey" a work means any kind of propagation that enables other 100 | parties to make or receive copies. Mere interaction with a user through 101 | a computer network, with no transfer of a copy, is not conveying. 102 | 103 | An interactive user interface displays "Appropriate Legal Notices" 104 | to the extent that it includes a convenient and prominently visible 105 | feature that (1) displays an appropriate copyright notice, and (2) 106 | tells the user that there is no warranty for the work (except to the 107 | extent that warranties are provided), that licensees may convey the 108 | work under this License, and how to view a copy of this License. If 109 | the interface presents a list of user commands or options, such as a 110 | menu, a prominent item in the list meets this criterion. 111 | 112 | 1. Source Code. 113 | 114 | The "source code" for a work means the preferred form of the work 115 | for making modifications to it. "Object code" means any non-source 116 | form of a work. 117 | 118 | A "Standard Interface" means an interface that either is an official 119 | standard defined by a recognized standards body, or, in the case of 120 | interfaces specified for a particular programming language, one that 121 | is widely used among developers working in that language. 122 | 123 | The "System Libraries" of an executable work include anything, other 124 | than the work as a whole, that (a) is included in the normal form of 125 | packaging a Major Component, but which is not part of that Major 126 | Component, and (b) serves only to enable use of the work with that 127 | Major Component, or to implement a Standard Interface for which an 128 | implementation is available to the public in source code form. A 129 | "Major Component", in this context, means a major essential component 130 | (kernel, window system, and so on) of the specific operating system 131 | (if any) on which the executable work runs, or a compiler used to 132 | produce the work, or an object code interpreter used to run it. 133 | 134 | The "Corresponding Source" for a work in object code form means all 135 | the source code needed to generate, install, and (for an executable 136 | work) run the object code and to modify the work, including scripts to 137 | control those activities. However, it does not include the work's 138 | System Libraries, or general-purpose tools or generally available free 139 | programs which are used unmodified in performing those activities but 140 | which are not part of the work. For example, Corresponding Source 141 | includes interface definition files associated with source files for 142 | the work, and the source code for shared libraries and dynamically 143 | linked subprograms that the work is specifically designed to require, 144 | such as by intimate data communication or control flow between those 145 | subprograms and other parts of the work. 146 | 147 | The Corresponding Source need not include anything that users 148 | can regenerate automatically from other parts of the Corresponding 149 | Source. 150 | 151 | The Corresponding Source for a work in source code form is that 152 | same work. 153 | 154 | 2. Basic Permissions. 155 | 156 | All rights granted under this License are granted for the term of 157 | copyright on the Program, and are irrevocable provided the stated 158 | conditions are met. This License explicitly affirms your unlimited 159 | permission to run the unmodified Program. The output from running a 160 | covered work is covered by this License only if the output, given its 161 | content, constitutes a covered work. This License acknowledges your 162 | rights of fair use or other equivalent, as provided by copyright law. 163 | 164 | You may make, run and propagate covered works that you do not 165 | convey, without conditions so long as your license otherwise remains 166 | in force. You may convey covered works to others for the sole purpose 167 | of having them make modifications exclusively for you, or provide you 168 | with facilities for running those works, provided that you comply with 169 | the terms of this License in conveying all material for which you do 170 | not control copyright. Those thus making or running the covered works 171 | for you must do so exclusively on your behalf, under your direction 172 | and control, on terms that prohibit them from making any copies of 173 | your copyrighted material outside their relationship with you. 174 | 175 | Conveying under any other circumstances is permitted solely under 176 | the conditions stated below. Sublicensing is not allowed; section 10 177 | makes it unnecessary. 178 | 179 | 3. Protecting Users' Legal Rights From Anti-Circumvention Law. 180 | 181 | No covered work shall be deemed part of an effective technological 182 | measure under any applicable law fulfilling obligations under article 183 | 11 of the WIPO copyright treaty adopted on 20 December 1996, or 184 | similar laws prohibiting or restricting circumvention of such 185 | measures. 186 | 187 | When you convey a covered work, you waive any legal power to forbid 188 | circumvention of technological measures to the extent such circumvention 189 | is effected by exercising rights under this License with respect to 190 | the covered work, and you disclaim any intention to limit operation or 191 | modification of the work as a means of enforcing, against the work's 192 | users, your or third parties' legal rights to forbid circumvention of 193 | technological measures. 194 | 195 | 4. Conveying Verbatim Copies. 196 | 197 | You may convey verbatim copies of the Program's source code as you 198 | receive it, in any medium, provided that you conspicuously and 199 | appropriately publish on each copy an appropriate copyright notice; 200 | keep intact all notices stating that this License and any 201 | non-permissive terms added in accord with section 7 apply to the code; 202 | keep intact all notices of the absence of any warranty; and give all 203 | recipients a copy of this License along with the Program. 204 | 205 | You may charge any price or no price for each copy that you convey, 206 | and you may offer support or warranty protection for a fee. 207 | 208 | 5. Conveying Modified Source Versions. 209 | 210 | You may convey a work based on the Program, or the modifications to 211 | produce it from the Program, in the form of source code under the 212 | terms of section 4, provided that you also meet all of these conditions: 213 | 214 | a) The work must carry prominent notices stating that you modified 215 | it, and giving a relevant date. 216 | 217 | b) The work must carry prominent notices stating that it is 218 | released under this License and any conditions added under section 219 | 7. This requirement modifies the requirement in section 4 to 220 | "keep intact all notices". 221 | 222 | c) You must license the entire work, as a whole, under this 223 | License to anyone who comes into possession of a copy. This 224 | License will therefore apply, along with any applicable section 7 225 | additional terms, to the whole of the work, and all its parts, 226 | regardless of how they are packaged. This License gives no 227 | permission to license the work in any other way, but it does not 228 | invalidate such permission if you have separately received it. 229 | 230 | d) If the work has interactive user interfaces, each must display 231 | Appropriate Legal Notices; however, if the Program has interactive 232 | interfaces that do not display Appropriate Legal Notices, your 233 | work need not make them do so. 234 | 235 | A compilation of a covered work with other separate and independent 236 | works, which are not by their nature extensions of the covered work, 237 | and which are not combined with it such as to form a larger program, 238 | in or on a volume of a storage or distribution medium, is called an 239 | "aggregate" if the compilation and its resulting copyright are not 240 | used to limit the access or legal rights of the compilation's users 241 | beyond what the individual works permit. Inclusion of a covered work 242 | in an aggregate does not cause this License to apply to the other 243 | parts of the aggregate. 244 | 245 | 6. Conveying Non-Source Forms. 246 | 247 | You may convey a covered work in object code form under the terms 248 | of sections 4 and 5, provided that you also convey the 249 | machine-readable Corresponding Source under the terms of this License, 250 | in one of these ways: 251 | 252 | a) Convey the object code in, or embodied in, a physical product 253 | (including a physical distribution medium), accompanied by the 254 | Corresponding Source fixed on a durable physical medium 255 | customarily used for software interchange. 256 | 257 | b) Convey the object code in, or embodied in, a physical product 258 | (including a physical distribution medium), accompanied by a 259 | written offer, valid for at least three years and valid for as 260 | long as you offer spare parts or customer support for that product 261 | model, to give anyone who possesses the object code either (1) a 262 | copy of the Corresponding Source for all the software in the 263 | product that is covered by this License, on a durable physical 264 | medium customarily used for software interchange, for a price no 265 | more than your reasonable cost of physically performing this 266 | conveying of source, or (2) access to copy the 267 | Corresponding Source from a network server at no charge. 268 | 269 | c) Convey individual copies of the object code with a copy of the 270 | written offer to provide the Corresponding Source. This 271 | alternative is allowed only occasionally and noncommercially, and 272 | only if you received the object code with such an offer, in accord 273 | with subsection 6b. 274 | 275 | d) Convey the object code by offering access from a designated 276 | place (gratis or for a charge), and offer equivalent access to the 277 | Corresponding Source in the same way through the same place at no 278 | further charge. You need not require recipients to copy the 279 | Corresponding Source along with the object code. If the place to 280 | copy the object code is a network server, the Corresponding Source 281 | may be on a different server (operated by you or a third party) 282 | that supports equivalent copying facilities, provided you maintain 283 | clear directions next to the object code saying where to find the 284 | Corresponding Source. Regardless of what server hosts the 285 | Corresponding Source, you remain obligated to ensure that it is 286 | available for as long as needed to satisfy these requirements. 287 | 288 | e) Convey the object code using peer-to-peer transmission, provided 289 | you inform other peers where the object code and Corresponding 290 | Source of the work are being offered to the general public at no 291 | charge under subsection 6d. 292 | 293 | A separable portion of the object code, whose source code is excluded 294 | from the Corresponding Source as a System Library, need not be 295 | included in conveying the object code work. 296 | 297 | A "User Product" is either (1) a "consumer product", which means any 298 | tangible personal property which is normally used for personal, family, 299 | or household purposes, or (2) anything designed or sold for incorporation 300 | into a dwelling. In determining whether a product is a consumer product, 301 | doubtful cases shall be resolved in favor of coverage. For a particular 302 | product received by a particular user, "normally used" refers to a 303 | typical or common use of that class of product, regardless of the status 304 | of the particular user or of the way in which the particular user 305 | actually uses, or expects or is expected to use, the product. A product 306 | is a consumer product regardless of whether the product has substantial 307 | commercial, industrial or non-consumer uses, unless such uses represent 308 | the only significant mode of use of the product. 309 | 310 | "Installation Information" for a User Product means any methods, 311 | procedures, authorization keys, or other information required to install 312 | and execute modified versions of a covered work in that User Product from 313 | a modified version of its Corresponding Source. The information must 314 | suffice to ensure that the continued functioning of the modified object 315 | code is in no case prevented or interfered with solely because 316 | modification has been made. 317 | 318 | If you convey an object code work under this section in, or with, or 319 | specifically for use in, a User Product, and the conveying occurs as 320 | part of a transaction in which the right of possession and use of the 321 | User Product is transferred to the recipient in perpetuity or for a 322 | fixed term (regardless of how the transaction is characterized), the 323 | Corresponding Source conveyed under this section must be accompanied 324 | by the Installation Information. But this requirement does not apply 325 | if neither you nor any third party retains the ability to install 326 | modified object code on the User Product (for example, the work has 327 | been installed in ROM). 328 | 329 | The requirement to provide Installation Information does not include a 330 | requirement to continue to provide support service, warranty, or updates 331 | for a work that has been modified or installed by the recipient, or for 332 | the User Product in which it has been modified or installed. Access to a 333 | network may be denied when the modification itself materially and 334 | adversely affects the operation of the network or violates the rules and 335 | protocols for communication across the network. 336 | 337 | Corresponding Source conveyed, and Installation Information provided, 338 | in accord with this section must be in a format that is publicly 339 | documented (and with an implementation available to the public in 340 | source code form), and must require no special password or key for 341 | unpacking, reading or copying. 342 | 343 | 7. Additional Terms. 344 | 345 | "Additional permissions" are terms that supplement the terms of this 346 | License by making exceptions from one or more of its conditions. 347 | Additional permissions that are applicable to the entire Program shall 348 | be treated as though they were included in this License, to the extent 349 | that they are valid under applicable law. If additional permissions 350 | apply only to part of the Program, that part may be used separately 351 | under those permissions, but the entire Program remains governed by 352 | this License without regard to the additional permissions. 353 | 354 | When you convey a copy of a covered work, you may at your option 355 | remove any additional permissions from that copy, or from any part of 356 | it. (Additional permissions may be written to require their own 357 | removal in certain cases when you modify the work.) You may place 358 | additional permissions on material, added by you to a covered work, 359 | for which you have or can give appropriate copyright permission. 360 | 361 | Notwithstanding any other provision of this License, for material you 362 | add to a covered work, you may (if authorized by the copyright holders of 363 | that material) supplement the terms of this License with terms: 364 | 365 | a) Disclaiming warranty or limiting liability differently from the 366 | terms of sections 15 and 16 of this License; or 367 | 368 | b) Requiring preservation of specified reasonable legal notices or 369 | author attributions in that material or in the Appropriate Legal 370 | Notices displayed by works containing it; or 371 | 372 | c) Prohibiting misrepresentation of the origin of that material, or 373 | requiring that modified versions of such material be marked in 374 | reasonable ways as different from the original version; or 375 | 376 | d) Limiting the use for publicity purposes of names of licensors or 377 | authors of the material; or 378 | 379 | e) Declining to grant rights under trademark law for use of some 380 | trade names, trademarks, or service marks; or 381 | 382 | f) Requiring indemnification of licensors and authors of that 383 | material by anyone who conveys the material (or modified versions of 384 | it) with contractual assumptions of liability to the recipient, for 385 | any liability that these contractual assumptions directly impose on 386 | those licensors and authors. 387 | 388 | All other non-permissive additional terms are considered "further 389 | restrictions" within the meaning of section 10. If the Program as you 390 | received it, or any part of it, contains a notice stating that it is 391 | governed by this License along with a term that is a further 392 | restriction, you may remove that term. If a license document contains 393 | a further restriction but permits relicensing or conveying under this 394 | License, you may add to a covered work material governed by the terms 395 | of that license document, provided that the further restriction does 396 | not survive such relicensing or conveying. 397 | 398 | If you add terms to a covered work in accord with this section, you 399 | must place, in the relevant source files, a statement of the 400 | additional terms that apply to those files, or a notice indicating 401 | where to find the applicable terms. 402 | 403 | Additional terms, permissive or non-permissive, may be stated in the 404 | form of a separately written license, or stated as exceptions; 405 | the above requirements apply either way. 406 | 407 | 8. Termination. 408 | 409 | You may not propagate or modify a covered work except as expressly 410 | provided under this License. Any attempt otherwise to propagate or 411 | modify it is void, and will automatically terminate your rights under 412 | this License (including any patent licenses granted under the third 413 | paragraph of section 11). 414 | 415 | However, if you cease all violation of this License, then your 416 | license from a particular copyright holder is reinstated (a) 417 | provisionally, unless and until the copyright holder explicitly and 418 | finally terminates your license, and (b) permanently, if the copyright 419 | holder fails to notify you of the violation by some reasonable means 420 | prior to 60 days after the cessation. 421 | 422 | Moreover, your license from a particular copyright holder is 423 | reinstated permanently if the copyright holder notifies you of the 424 | violation by some reasonable means, this is the first time you have 425 | received notice of violation of this License (for any work) from that 426 | copyright holder, and you cure the violation prior to 30 days after 427 | your receipt of the notice. 428 | 429 | Termination of your rights under this section does not terminate the 430 | licenses of parties who have received copies or rights from you under 431 | this License. If your rights have been terminated and not permanently 432 | reinstated, you do not qualify to receive new licenses for the same 433 | material under section 10. 434 | 435 | 9. Acceptance Not Required for Having Copies. 436 | 437 | You are not required to accept this License in order to receive or 438 | run a copy of the Program. Ancillary propagation of a covered work 439 | occurring solely as a consequence of using peer-to-peer transmission 440 | to receive a copy likewise does not require acceptance. However, 441 | nothing other than this License grants you permission to propagate or 442 | modify any covered work. These actions infringe copyright if you do 443 | not accept this License. Therefore, by modifying or propagating a 444 | covered work, you indicate your acceptance of this License to do so. 445 | 446 | 10. Automatic Licensing of Downstream Recipients. 447 | 448 | Each time you convey a covered work, the recipient automatically 449 | receives a license from the original licensors, to run, modify and 450 | propagate that work, subject to this License. You are not responsible 451 | for enforcing compliance by third parties with this License. 452 | 453 | An "entity transaction" is a transaction transferring control of an 454 | organization, or substantially all assets of one, or subdividing an 455 | organization, or merging organizations. If propagation of a covered 456 | work results from an entity transaction, each party to that 457 | transaction who receives a copy of the work also receives whatever 458 | licenses to the work the party's predecessor in interest had or could 459 | give under the previous paragraph, plus a right to possession of the 460 | Corresponding Source of the work from the predecessor in interest, if 461 | the predecessor has it or can get it with reasonable efforts. 462 | 463 | You may not impose any further restrictions on the exercise of the 464 | rights granted or affirmed under this License. For example, you may 465 | not impose a license fee, royalty, or other charge for exercise of 466 | rights granted under this License, and you may not initiate litigation 467 | (including a cross-claim or counterclaim in a lawsuit) alleging that 468 | any patent claim is infringed by making, using, selling, offering for 469 | sale, or importing the Program or any portion of it. 470 | 471 | 11. Patents. 472 | 473 | A "contributor" is a copyright holder who authorizes use under this 474 | License of the Program or a work on which the Program is based. The 475 | work thus licensed is called the contributor's "contributor version". 476 | 477 | A contributor's "essential patent claims" are all patent claims 478 | owned or controlled by the contributor, whether already acquired or 479 | hereafter acquired, that would be infringed by some manner, permitted 480 | by this License, of making, using, or selling its contributor version, 481 | but do not include claims that would be infringed only as a 482 | consequence of further modification of the contributor version. For 483 | purposes of this definition, "control" includes the right to grant 484 | patent sublicenses in a manner consistent with the requirements of 485 | this License. 486 | 487 | Each contributor grants you a non-exclusive, worldwide, royalty-free 488 | patent license under the contributor's essential patent claims, to 489 | make, use, sell, offer for sale, import and otherwise run, modify and 490 | propagate the contents of its contributor version. 491 | 492 | In the following three paragraphs, a "patent license" is any express 493 | agreement or commitment, however denominated, not to enforce a patent 494 | (such as an express permission to practice a patent or covenant not to 495 | sue for patent infringement). To "grant" such a patent license to a 496 | party means to make such an agreement or commitment not to enforce a 497 | patent against the party. 498 | 499 | If you convey a covered work, knowingly relying on a patent license, 500 | and the Corresponding Source of the work is not available for anyone 501 | to copy, free of charge and under the terms of this License, through a 502 | publicly available network server or other readily accessible means, 503 | then you must either (1) cause the Corresponding Source to be so 504 | available, or (2) arrange to deprive yourself of the benefit of the 505 | patent license for this particular work, or (3) arrange, in a manner 506 | consistent with the requirements of this License, to extend the patent 507 | license to downstream recipients. "Knowingly relying" means you have 508 | actual knowledge that, but for the patent license, your conveying the 509 | covered work in a country, or your recipient's use of the covered work 510 | in a country, would infringe one or more identifiable patents in that 511 | country that you have reason to believe are valid. 512 | 513 | If, pursuant to or in connection with a single transaction or 514 | arrangement, you convey, or propagate by procuring conveyance of, a 515 | covered work, and grant a patent license to some of the parties 516 | receiving the covered work authorizing them to use, propagate, modify 517 | or convey a specific copy of the covered work, then the patent license 518 | you grant is automatically extended to all recipients of the covered 519 | work and works based on it. 520 | 521 | A patent license is "discriminatory" if it does not include within 522 | the scope of its coverage, prohibits the exercise of, or is 523 | conditioned on the non-exercise of one or more of the rights that are 524 | specifically granted under this License. You may not convey a covered 525 | work if you are a party to an arrangement with a third party that is 526 | in the business of distributing software, under which you make payment 527 | to the third party based on the extent of your activity of conveying 528 | the work, and under which the third party grants, to any of the 529 | parties who would receive the covered work from you, a discriminatory 530 | patent license (a) in connection with copies of the covered work 531 | conveyed by you (or copies made from those copies), or (b) primarily 532 | for and in connection with specific products or compilations that 533 | contain the covered work, unless you entered into that arrangement, 534 | or that patent license was granted, prior to 28 March 2007. 535 | 536 | Nothing in this License shall be construed as excluding or limiting 537 | any implied license or other defenses to infringement that may 538 | otherwise be available to you under applicable patent law. 539 | 540 | 12. No Surrender of Others' Freedom. 541 | 542 | If conditions are imposed on you (whether by court order, agreement or 543 | otherwise) that contradict the conditions of this License, they do not 544 | excuse you from the conditions of this License. If you cannot convey a 545 | covered work so as to satisfy simultaneously your obligations under this 546 | License and any other pertinent obligations, then as a consequence you may 547 | not convey it at all. For example, if you agree to terms that obligate you 548 | to collect a royalty for further conveying from those to whom you convey 549 | the Program, the only way you could satisfy both those terms and this 550 | License would be to refrain entirely from conveying the Program. 551 | 552 | 13. Use with the GNU Affero General Public License. 553 | 554 | Notwithstanding any other provision of this License, you have 555 | permission to link or combine any covered work with a work licensed 556 | under version 3 of the GNU Affero General Public License into a single 557 | combined work, and to convey the resulting work. The terms of this 558 | License will continue to apply to the part which is the covered work, 559 | but the special requirements of the GNU Affero General Public License, 560 | section 13, concerning interaction through a network will apply to the 561 | combination as such. 562 | 563 | 14. Revised Versions of this License. 564 | 565 | The Free Software Foundation may publish revised and/or new versions of 566 | the GNU General Public License from time to time. Such new versions will 567 | be similar in spirit to the present version, but may differ in detail to 568 | address new problems or concerns. 569 | 570 | Each version is given a distinguishing version number. If the 571 | Program specifies that a certain numbered version of the GNU General 572 | Public License "or any later version" applies to it, you have the 573 | option of following the terms and conditions either of that numbered 574 | version or of any later version published by the Free Software 575 | Foundation. If the Program does not specify a version number of the 576 | GNU General Public License, you may choose any version ever published 577 | by the Free Software Foundation. 578 | 579 | If the Program specifies that a proxy can decide which future 580 | versions of the GNU General Public License can be used, that proxy's 581 | public statement of acceptance of a version permanently authorizes you 582 | to choose that version for the Program. 583 | 584 | Later license versions may give you additional or different 585 | permissions. However, no additional obligations are imposed on any 586 | author or copyright holder as a result of your choosing to follow a 587 | later version. 588 | 589 | 15. Disclaimer of Warranty. 590 | 591 | THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY 592 | APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT 593 | HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY 594 | OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO, 595 | THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 596 | PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM 597 | IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF 598 | ALL NECESSARY SERVICING, REPAIR OR CORRECTION. 599 | 600 | 16. Limitation of Liability. 601 | 602 | IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING 603 | WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS 604 | THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY 605 | GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE 606 | USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF 607 | DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD 608 | PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS), 609 | EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF 610 | SUCH DAMAGES. 611 | 612 | 17. Interpretation of Sections 15 and 16. 613 | 614 | If the disclaimer of warranty and limitation of liability provided 615 | above cannot be given local legal effect according to their terms, 616 | reviewing courts shall apply local law that most closely approximates 617 | an absolute waiver of all civil liability in connection with the 618 | Program, unless a warranty or assumption of liability accompanies a 619 | copy of the Program in return for a fee. 620 | 621 | END OF TERMS AND CONDITIONS 622 | 623 | How to Apply These Terms to Your New Programs 624 | 625 | If you develop a new program, and you want it to be of the greatest 626 | possible use to the public, the best way to achieve this is to make it 627 | free software which everyone can redistribute and change under these terms. 628 | 629 | To do so, attach the following notices to the program. It is safest 630 | to attach them to the start of each source file to most effectively 631 | state the exclusion of warranty; and each file should have at least 632 | the "copyright" line and a pointer to where the full notice is found. 633 | 634 | 635 | Copyright (C) 636 | 637 | This program is free software: you can redistribute it and/or modify 638 | it under the terms of the GNU General Public License as published by 639 | the Free Software Foundation, either version 3 of the License, or 640 | (at your option) any later version. 641 | 642 | This program is distributed in the hope that it will be useful, 643 | but WITHOUT ANY WARRANTY; without even the implied warranty of 644 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 645 | GNU General Public License for more details. 646 | 647 | You should have received a copy of the GNU General Public License 648 | along with this program. If not, see . 649 | 650 | Also add information on how to contact you by electronic and paper mail. 651 | 652 | If the program does terminal interaction, make it output a short 653 | notice like this when it starts in an interactive mode: 654 | 655 | Copyright (C) 656 | This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'. 657 | This is free software, and you are welcome to redistribute it 658 | under certain conditions; type `show c' for details. 659 | 660 | The hypothetical commands `show w' and `show c' should show the appropriate 661 | parts of the General Public License. Of course, your program's commands 662 | might be different; for a GUI interface, you would use an "about box". 663 | 664 | You should also get your employer (if you work as a programmer) or school, 665 | if any, to sign a "copyright disclaimer" for the program, if necessary. 666 | For more information on this, and how to apply and follow the GNU GPL, see 667 | . 668 | 669 | The GNU General Public License does not permit incorporating your program 670 | into proprietary programs. If your program is a subroutine library, you 671 | may consider it more useful to permit linking proprietary applications with 672 | the library. If this is what you want to do, use the GNU Lesser General 673 | Public License instead of this License. But first, please read 674 | . 675 | -------------------------------------------------------------------------------- /Supplymental_material.pdf: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/nagejacob/SpatiallyAdaptiveSSID/7e94712f2b38fd926b0b6e879e9d85f09e942cc4/Supplymental_material.pdf -------------------------------------------------------------------------------- /dataset/DND_sRGB.py: -------------------------------------------------------------------------------- 1 | from dataset.base import dataset_path 2 | import h5py 3 | import glob 4 | import numpy as np 5 | import os 6 | import scipy.io as sio 7 | from torch.utils.data import Dataset 8 | 9 | dnd_path = os.path.join(dataset_path, 'DND') 10 | 11 | class DNDSrgbBenchmarkDataset(Dataset): 12 | def __init__(self): 13 | super(DNDSrgbBenchmarkDataset, self).__init__() 14 | self.imgs = [] 15 | infos = h5py.File(os.path.join(dnd_path, 'info.mat'), 'r') 16 | info = infos['info'] 17 | bb = info['boundingboxes'] 18 | for i in range(50): 19 | filename = os.path.join(dnd_path, 'images_srgb', '%04d.mat' % (i + 1)) 20 | img = h5py.File(filename, 'r') 21 | Inoisy = np.float32(np.array(img['InoisySRGB']).T) 22 | ref = bb[0][i] 23 | boxes = np.array(info[ref]).T 24 | for k in range(20): 25 | idx = [int(boxes[k, 0] - 1), int(boxes[k, 2]), int(boxes[k, 1] - 1), int(boxes[k, 3])] 26 | Inoisy_crop = Inoisy[idx[0]:idx[1], idx[2]:idx[3], :].copy() 27 | Inoisy_crop = np.transpose(Inoisy_crop, (2, 0, 1)) * 255. 28 | 29 | self.imgs.append({'L':Inoisy_crop}) 30 | 31 | def __getitem__(self, index): 32 | return self.imgs[index] 33 | 34 | def __len__(self): 35 | return 1000 -------------------------------------------------------------------------------- /dataset/SIDD_sRGB.py: -------------------------------------------------------------------------------- 1 | ''' 2 | We observe slightly better performance with training inputs in [0, 255] range than that in [0, 1], 3 | so we follow AP-BSN that do not normalize the input image from [0, 255] to [0, 1]. 4 | ''' 5 | from dataset.base import BaseTrainDataset, dataset_path 6 | import glob 7 | import numpy as np 8 | import os 9 | from PIL import Image 10 | import scipy.io as sio 11 | from torch.utils.data import Dataset 12 | 13 | sidd_path = os.path.join(dataset_path, 'SIDD') 14 | 15 | class SIDDSrgbTrainDataset(BaseTrainDataset): 16 | def __init__(self, patch_size, pin_memory): 17 | super(SIDDSrgbTrainDataset, self).__init__(sidd_path, patch_size, pin_memory) 18 | 19 | def __getitem__(self, index): 20 | index = index % len(self.img_paths) 21 | 22 | if self.pin_memory: 23 | img_L = self.imgs[index]['L'] 24 | img_H = self.imgs[index]['H'] 25 | else: 26 | img_path = self.img_paths[index] 27 | img_L = self._open_image(img_path['L']) 28 | img_H = self._open_image(img_path['H']) 29 | 30 | img_L, img_H = self.crop(img_L, img_H) 31 | img_L, img_H = self.augment(img_L, img_H) 32 | 33 | img_L, img_H = np.float32(img_L), np.float32(img_H) 34 | return {'L': img_L, 'H': img_H} 35 | 36 | def _get_img_paths(self, path): 37 | self.img_paths = [] 38 | L_pattern = os.path.join(path, 'SIDD_Medium_Srgb/Data/*/*_NOISY_SRGB_*.PNG') 39 | L_paths = sorted(glob.glob(L_pattern)) 40 | for L_path in L_paths: 41 | self.img_paths.append({'L': L_path, 'H': L_path.replace('NOISY', 'GT')}) 42 | 43 | def _open_images(self): 44 | self.imgs = [] 45 | for img_path in self.img_paths: 46 | img_L = self._open_image(img_path['L']) 47 | img_H = self._open_image(img_path['H']) 48 | self.imgs.append({'L': img_L, 'H': img_H}) 49 | 50 | def _open_image(self, path): 51 | img = Image.open(path) 52 | img = np.asarray(img) 53 | img = np.transpose(img, (2, 0, 1)) 54 | return img 55 | 56 | 57 | class SIDDSrgbValidationDataset(Dataset): 58 | def __init__(self): 59 | super(SIDDSrgbValidationDataset, self).__init__() 60 | self._open_images(sidd_path) 61 | self.n = self.noisy_block.shape[0] 62 | self.k = self.noisy_block.shape[1] 63 | 64 | def __getitem__(self, index): 65 | index_n = index // self.k 66 | index_k = index % self.k 67 | 68 | img_H = self.gt_block[index_n, index_k] 69 | img_H = np.float32(img_H) 70 | img_H = np.transpose(img_H, (2, 0, 1)) 71 | 72 | img_L = self.noisy_block[index_n, index_k] 73 | img_L = np.float32(img_L) 74 | img_L = np.transpose(img_L, (2, 0, 1)) 75 | 76 | return {'H':img_H, 'L':img_L} 77 | 78 | def __len__(self): 79 | return self.n * self.k 80 | 81 | def _open_images(self, path): 82 | mat = sio.loadmat(os.path.join(path, 'SIDD_Validation/ValidationNoisyBlocksSrgb.mat')) 83 | self.noisy_block = mat['ValidationNoisyBlocksSrgb'] 84 | mat = sio.loadmat(os.path.join(path, 'SIDD_Validation/ValidationGtBlocksSrgb.mat')) 85 | self.gt_block = mat['ValidationGtBlocksSrgb'] 86 | 87 | 88 | class SIDDSrgbBenchmarkDataset(Dataset): 89 | def __init__(self): 90 | super(SIDDSrgbBenchmarkDataset, self).__init__() 91 | self._open_images(sidd_path) 92 | self.n = self.noisy_block.shape[0] 93 | self.k = self.noisy_block.shape[1] 94 | 95 | def __getitem__(self, index): 96 | index_n = index // self.k 97 | index_k = index % self.k 98 | 99 | img_L = self.noisy_block[index_n, index_k] 100 | img_L = np.float32(img_L) 101 | img_L = np.transpose(img_L, (2, 0, 1)) 102 | 103 | return {'L':img_L} 104 | 105 | def __len__(self): 106 | return self.n * self.k 107 | 108 | def _open_images(self, path): 109 | mat = sio.loadmat(os.path.join(path, 'SIDD_Benchmark/BenchmarkNoisyBlocksSrgb.mat')) 110 | self.noisy_block = mat['BenchmarkNoisyBlocksSrgb'] -------------------------------------------------------------------------------- /dataset/__init__.py: -------------------------------------------------------------------------------- 1 | from dataset.SIDD_sRGB import SIDDSrgbTrainDataset, SIDDSrgbValidationDataset, SIDDSrgbBenchmarkDataset 2 | from dataset.DND_sRGB import DNDSrgbBenchmarkDataset -------------------------------------------------------------------------------- /dataset/base.py: -------------------------------------------------------------------------------- 1 | from abc import abstractmethod 2 | import numpy as np 3 | import random 4 | import socket 5 | import torch.utils.data as data 6 | 7 | hostname = socket.gethostname() 8 | if 'lijunyis-ubuntu' == hostname: 9 | dataset_path = '/home/nagejacob/Documents/datasets' 10 | else: 11 | raise OSError # dataset_path = 'path_to_dataset' 12 | 13 | # c, h, w numpy 14 | def aug_np3(img, flip_h, flip_w, transpose): 15 | if flip_h: 16 | img = img[:, ::-1, :] 17 | if flip_w: 18 | img = img[:, :, ::-1] 19 | if transpose: 20 | img = np.transpose(img, (0, 2, 1)) 21 | 22 | return img 23 | 24 | def crop_np3(img, patch_size, position_h, position_w): 25 | return img[:, position_h:position_h+patch_size, position_w:position_w+patch_size] 26 | 27 | class BaseTrainDataset(data.Dataset): 28 | def __init__(self, path, patch_size, pin_memory): 29 | super(BaseTrainDataset, self).__init__() 30 | self.patch_size = patch_size 31 | self.pin_memory = pin_memory 32 | self._get_img_paths(path) 33 | if self.pin_memory: 34 | self._open_images() 35 | 36 | @abstractmethod 37 | def __getitem__(self, index): 38 | pass 39 | 40 | def __len__(self): 41 | return 100000 42 | 43 | @abstractmethod 44 | def _get_img_paths(self, path): 45 | pass 46 | 47 | @abstractmethod 48 | def _open_images(self): 49 | pass 50 | 51 | @abstractmethod 52 | def _open_image(self, path): 53 | pass 54 | 55 | def crop(self, img_L, img_H=None): 56 | C, H, W = img_L.shape 57 | position_H = random.randint(0, H - self.patch_size) 58 | position_W = random.randint(0, W - self.patch_size) 59 | 60 | patch_L = crop_np3(img_L, self.patch_size, position_H, position_W) 61 | if img_H is not None: 62 | patch_H = crop_np3(img_H, self.patch_size, position_H, position_W) 63 | return patch_L, patch_H 64 | else: 65 | return patch_L 66 | 67 | def augment(self, img_L, img_H=None): 68 | flip_h = random.random() > 0.5 69 | flip_w = random.random() > 0.5 70 | transpose = random.random() > 0.5 71 | img_L = aug_np3(img_L, flip_h, flip_w, transpose) 72 | if img_H is not None: 73 | img_H = aug_np3(img_H, flip_h, flip_w, transpose) 74 | return img_L, img_H 75 | else: 76 | return img_L -------------------------------------------------------------------------------- /image_denoising.yaml: -------------------------------------------------------------------------------- 1 | name: /hdd/Documents/anaconda/image_denoising 2 | channels: 3 | - https://mirrors.hit.edu.cn/anaconda/cloud/pytorch 4 | - anaconda 5 | - https://mirrors.hit.edu.cn/anaconda/cloud/conda-forge 6 | - defaults 7 | dependencies: 8 | - _libgcc_mutex=0.1=main 9 | - _openmp_mutex=4.5=1_gnu 10 | - absl-py=1.0.0=pyhd8ed1ab_0 11 | - aiohttp=3.7.4.post0=py38h7f8727e_2 12 | - alsa-lib=1.2.3=h516909a_0 13 | - appdirs=1.4.4=pyh9f0ad1d_0 14 | - argon2-cffi=21.3.0=pyhd8ed1ab_0 15 | - argon2-cffi-bindings=21.2.0=py38h0a891b7_2 16 | - asttokens=2.0.8=pyhd8ed1ab_0 17 | - async-timeout=3.0.1=py_1000 18 | - attrs=21.4.0=pyhd8ed1ab_0 19 | - backcall=0.2.0=pyh9f0ad1d_0 20 | - backports=1.0=py_2 21 | - backports.functools_lru_cache=1.6.4=pyhd8ed1ab_0 22 | - beautifulsoup4=4.11.1=pyha770c72_0 23 | - blas=1.0=mkl 24 | - bleach=5.0.1=pyhd8ed1ab_0 25 | - blinker=1.4=py_1 26 | - blosc=1.21.0=h8c45485_0 27 | - bottleneck=1.3.4=py38hce1f21e_0 28 | - brotli=1.0.9=he6710b0_2 29 | - brotlipy=0.7.0=py38h27cfd23_1003 30 | - brunsli=0.1=h2531618_0 31 | - bzip2=1.0.8=h7b6447c_0 32 | - c-ares=1.17.1=h7f98852_1 33 | - ca-certificates=2022.12.7=ha878542_0 34 | - cachetools=5.0.0=pyhd8ed1ab_0 35 | - cairo=1.16.0=h6cf1ce9_1008 36 | - certifi=2022.12.7=pyhd8ed1ab_0 37 | - cffi=1.15.0=py38hd667e15_1 38 | - cfitsio=3.470=hf0d0db6_6 39 | - chardet=4.0.0=py38h578d9bd_3 40 | - charls=2.2.0=h2531618_0 41 | - charset-normalizer=2.0.4=pyhd3eb1b0_0 42 | - click=8.1.2=py38h578d9bd_0 43 | - cloudpickle=2.0.0=pyhd3eb1b0_0 44 | - colorama=0.4.4=pyhd3eb1b0_0 45 | - cryptography=36.0.0=py38h9ce1e76_0 46 | - cudatoolkit=11.3.1=h2bc3f7f_2 47 | - cycler=0.11.0=pyhd8ed1ab_0 48 | - cython=0.29.28=py38h295c915_0 49 | - cytoolz=0.11.0=py38h7b6447c_0 50 | - dask-core=2022.2.1=pyhd3eb1b0_0 51 | - dbus=1.13.6=h48d8840_2 52 | - decorator=5.1.1=pyhd8ed1ab_0 53 | - defusedxml=0.7.1=pyhd8ed1ab_0 54 | - docker-pycreds=0.4.0=py_0 55 | - einops=0.6.0=pyhd8ed1ab_0 56 | - entrypoints=0.4=pyhd8ed1ab_0 57 | - executing=1.1.0=pyhd8ed1ab_0 58 | - exifread=3.0.0=pyhd8ed1ab_0 59 | - expat=2.4.1=h9c3ff4c_0 60 | - ffmpeg=4.3=hf484d3e_0 61 | - flit-core=3.7.1=pyhd8ed1ab_0 62 | - fontconfig=2.13.1=hba837de_1005 63 | - fonttools=4.25.0=pyhd3eb1b0_0 64 | - freetype=2.11.0=h70c0345_0 65 | - fsspec=2022.2.0=pyhd3eb1b0_0 66 | - gettext=0.19.8.1=h0b5b191_1005 67 | - giflib=5.2.1=h7b6447c_0 68 | - gitdb=4.0.10=pyhd8ed1ab_0 69 | - gitpython=3.1.31=pyhd8ed1ab_0 70 | - glib=2.68.3=h9c3ff4c_0 71 | - glib-tools=2.68.3=h9c3ff4c_0 72 | - gmp=6.2.1=h2531618_2 73 | - gnutls=3.6.15=he1e5248_0 74 | - google-auth=2.6.5=pyh6c4a22f_0 75 | - google-auth-oauthlib=0.4.6=pyhd8ed1ab_0 76 | - graphite2=1.3.13=h58526e2_1001 77 | - grpcio=1.42.0=py38hce63b2e_0 78 | - gst-plugins-base=1.18.4=hf529b03_2 79 | - gstreamer=1.18.4=h76c114f_2 80 | - h5py=3.6.0=py38ha0f2276_0 81 | - harfbuzz=2.8.2=h83ec7ef_0 82 | - hdf5=1.10.6=hb1b8bf9_0 83 | - icu=68.1=h58526e2_0 84 | - idna=3.3=pyhd3eb1b0_0 85 | - imagecodecs=2021.8.26=py38h4cda21f_0 86 | - imageio=2.9.0=pyhd3eb1b0_0 87 | - importlib-metadata=4.11.3=py38h578d9bd_1 88 | - importlib_resources=5.9.0=pyhd8ed1ab_0 89 | - intel-openmp=2021.4.0=h06a4308_3561 90 | - ipykernel=5.5.5=py38hd0cf306_0 91 | - ipython=8.5.0=pyh41d4057_1 92 | - ipython_genutils=0.2.0=py_1 93 | - ipywidgets=8.0.2=pyhd8ed1ab_1 94 | - jasper=1.900.1=h07fcdf6_1006 95 | - jedi=0.18.1=pyhd8ed1ab_2 96 | - jinja2=3.1.2=pyhd8ed1ab_1 97 | - joblib=1.1.0=pyhd3eb1b0_0 98 | - jpeg=9d=h7f8727e_0 99 | - jsonschema=4.16.0=pyhd8ed1ab_0 100 | - jupyter=1.0.0=py38h578d9bd_7 101 | - jupyter_client=7.0.6=pyhd8ed1ab_0 102 | - jupyter_console=6.4.4=pyhd8ed1ab_0 103 | - jupyter_core=4.11.1=py38h578d9bd_0 104 | - jupyterlab_pygments=0.2.2=pyhd8ed1ab_0 105 | - jupyterlab_widgets=3.0.3=pyhd8ed1ab_0 106 | - jxrlib=1.1=h7b6447c_2 107 | - kiwisolver=1.4.2=py38h295c915_0 108 | - krb5=1.19.2=hcc1bbae_0 109 | - lame=3.100=h7b6447c_0 110 | - lcms2=2.12=h3be6417_0 111 | - ld_impl_linux-64=2.35.1=h7274673_9 112 | - lerc=3.0=h295c915_0 113 | - libaec=1.0.4=he6710b0_1 114 | - libblas=3.9.0=12_linux64_mkl 115 | - libcblas=3.9.0=12_linux64_mkl 116 | - libcurl=7.78.0=h2574ce0_0 117 | - libdeflate=1.8=h7f8727e_5 118 | - libedit=3.1.20191231=he28a2e2_2 119 | - libev=4.33=h516909a_1 120 | - libevent=2.1.10=hcdb4288_3 121 | - libffi=3.3=he6710b0_2 122 | - libgcc-ng=11.2.0=h1234567_0 123 | - libgfortran-ng=7.5.0=ha8ba4b0_17 124 | - libgfortran4=7.5.0=ha8ba4b0_17 125 | - libglib=2.68.3=h3e27bee_0 126 | - libgomp=11.2.0=h1234567_0 127 | - libiconv=1.16=h516909a_0 128 | - libidn2=2.3.2=h7f8727e_0 129 | - liblapack=3.9.0=12_linux64_mkl 130 | - liblapacke=3.9.0=12_linux64_mkl 131 | - libllvm11=11.1.0=hf817b99_2 132 | - libnghttp2=1.43.0=h812cca2_0 133 | - libogg=1.3.4=h7f98852_1 134 | - libopencv=4.4.0=py38_2 135 | - libopus=1.3.1=h7f98852_1 136 | - libpng=1.6.37=hbc83047_0 137 | - libpq=13.3=hd57d9b9_0 138 | - libprotobuf=3.20.3=he621ea3_0 139 | - libsodium=1.0.18=h36c2ea0_1 140 | - libssh2=1.9.0=ha56f1ee_6 141 | - libstdcxx-ng=12.1.0=ha89aaad_16 142 | - libtasn1=4.16.0=h27cfd23_0 143 | - libtiff=4.2.0=h85742a9_0 144 | - libunistring=0.9.10=h27cfd23_0 145 | - libuuid=2.32.1=h7f98852_1000 146 | - libuv=1.40.0=h7b6447c_0 147 | - libvorbis=1.3.7=h9c3ff4c_0 148 | - libwebp=1.2.2=h55f646e_0 149 | - libwebp-base=1.2.2=h7f8727e_0 150 | - libxcb=1.13=h7f98852_1003 151 | - libxkbcommon=1.0.3=he3ba5ed_0 152 | - libxml2=2.9.12=h72842e0_0 153 | - libzopfli=1.0.3=he6710b0_0 154 | - locket=0.2.1=py38h06a4308_2 155 | - lpips=0.1.3=pyhd8ed1ab_0 156 | - lz4-c=1.9.3=h295c915_1 157 | - mako=1.2.0=pyhd8ed1ab_1 158 | - markdown=3.3.6=pyhd8ed1ab_0 159 | - markupsafe=2.1.1=py38h0a891b7_1 160 | - matplotlib=3.5.1=py38h06a4308_1 161 | - matplotlib-base=3.5.1=py38ha18d171_1 162 | - matplotlib-inline=0.1.6=pyhd8ed1ab_0 163 | - mistune=0.8.4=py38h497a2fe_1005 164 | - mkl=2021.4.0=h06a4308_640 165 | - mkl-service=2.4.0=py38h7f8727e_0 166 | - mkl_fft=1.3.1=py38hd3c417c_0 167 | - mkl_random=1.2.2=py38h51133e4_0 168 | - multidict=5.2.0=py38h7f8727e_2 169 | - munkres=1.1.4=pyh9f0ad1d_0 170 | - mysql-common=8.0.25=ha770c72_0 171 | - mysql-libs=8.0.25=h935591d_0 172 | - natsort=7.1.1=pyhd3eb1b0_0 173 | - nbclient=0.5.13=pyhd8ed1ab_0 174 | - nbconvert=6.4.5=py38h578d9bd_0 175 | - nbformat=5.6.1=pyhd8ed1ab_0 176 | - ncurses=6.3=h7f8727e_2 177 | - nest-asyncio=1.5.5=pyhd8ed1ab_0 178 | - nettle=3.7.3=hbbd107a_1 179 | - networkx=2.7.1=pyhd3eb1b0_0 180 | - notebook=6.4.12=pyha770c72_0 181 | - nspr=4.30=h9c3ff4c_0 182 | - nss=3.67=hb5efdd6_0 183 | - numexpr=2.8.1=py38h807cd23_2 184 | - numpy=1.21.5=py38he7a7128_1 185 | - numpy-base=1.21.5=py38hf524024_1 186 | - oauthlib=3.2.0=pyhd8ed1ab_0 187 | - ocl-icd=2.3.1=h7f98852_0 188 | - ocl-icd-system=1.0.0=1 189 | - opencv=4.4.0=py38_2 190 | - openh264=2.1.1=h4ff587b_0 191 | - openjpeg=2.4.0=h3ad879b_0 192 | - openssl=1.1.1t=h7f8727e_0 193 | - packaging=21.3=pyhd3eb1b0_0 194 | - pandas=1.4.2=py38h295c915_0 195 | - pandoc=2.19.2=ha770c72_0 196 | - pandocfilters=1.5.0=pyhd8ed1ab_0 197 | - parso=0.8.3=pyhd8ed1ab_0 198 | - partd=1.2.0=pyhd3eb1b0_1 199 | - pathtools=0.1.2=py_1 200 | - pcre=8.45=h9c3ff4c_0 201 | - pexpect=4.8.0=pyh9f0ad1d_2 202 | - pickleshare=0.7.5=py_1003 203 | - pillow=9.0.1=py38h22f2fdc_0 204 | - pixman=0.40.0=h36c2ea0_0 205 | - pkgutil-resolve-name=1.3.10=pyhd8ed1ab_0 206 | - platformdirs=2.5.1=pyhd8ed1ab_0 207 | - prometheus_client=0.14.1=pyhd8ed1ab_0 208 | - prompt-toolkit=3.0.31=pyha770c72_0 209 | - prompt_toolkit=3.0.31=hd8ed1ab_0 210 | - protobuf=3.20.3=py38h6a678d5_0 211 | - psutil=5.9.0=py38h5eee18b_0 212 | - pthread-stubs=0.4=h36c2ea0_1001 213 | - ptyprocess=0.7.0=pyhd3deb0d_0 214 | - pure_eval=0.2.2=pyhd8ed1ab_0 215 | - py-opencv=4.4.0=py38h23f93f0_2 216 | - pyasn1=0.4.8=py_0 217 | - pyasn1-modules=0.2.7=py_0 218 | - pycparser=2.21=pyhd3eb1b0_0 219 | - pygments=2.13.0=pyhd8ed1ab_0 220 | - pyjwt=2.3.0=pyhd8ed1ab_1 221 | - pynvml=11.4.1=pyhd8ed1ab_0 222 | - pyopencl=2022.1.5=py38h514daf8_0 223 | - pyopenssl=22.0.0=pyhd3eb1b0_0 224 | - pyparsing=3.0.4=pyhd3eb1b0_0 225 | - pyqt=5.12.3=py38ha8c2ead_4 226 | - pyrsistent=0.18.1=py38h0a891b7_1 227 | - pysocks=1.7.1=py38h06a4308_0 228 | - python=3.8.13=h12debd9_0 229 | - python-dateutil=2.8.2=pyhd8ed1ab_0 230 | - python-fastjsonschema=2.16.2=pyhd8ed1ab_0 231 | - python_abi=3.8=2_cp38 232 | - pytools=2022.1.9=pyhd8ed1ab_0 233 | - pytorch=1.11.0=py3.8_cuda11.3_cudnn8.2.0_0 234 | - pytorch-mutex=1.0=cuda 235 | - pytz=2022.1=py38h06a4308_0 236 | - pyu2f=0.1.5=pyhd8ed1ab_0 237 | - pywavelets=1.3.0=py38h7f8727e_0 238 | - pyzmq=19.0.2=py38ha71036d_2 239 | - qt=5.12.9=hda022c4_4 240 | - qtconsole=5.3.2=pyhd8ed1ab_0 241 | - qtconsole-base=5.3.2=pyha770c72_0 242 | - qtpy=2.2.0=pyhd8ed1ab_0 243 | - readline=8.1.2=h7f8727e_1 244 | - requests=2.27.1=pyhd3eb1b0_0 245 | - requests-oauthlib=1.3.1=pyhd8ed1ab_0 246 | - rsa=4.8=pyhd8ed1ab_0 247 | - scikit-image=0.19.2=py38h51133e4_0 248 | - scikit-learn=1.1.1=py38h6a678d5_0 249 | - scipy=1.7.3=py38hc147768_0 250 | - send2trash=1.8.0=pyhd8ed1ab_0 251 | - sentry-sdk=1.19.1=pyhd8ed1ab_0 252 | - setproctitle=1.2.2=py38h0a891b7_2 253 | - setuptools=61.2.0=py38h06a4308_0 254 | - six=1.16.0=pyhd3eb1b0_1 255 | - smmap=3.0.5=pyh44b312d_0 256 | - snappy=1.1.8=he6710b0_0 257 | - soupsieve=2.3.2.post1=pyhd8ed1ab_0 258 | - sqlite=3.38.2=hc218d9a_0 259 | - stack_data=0.5.1=pyhd8ed1ab_0 260 | - tensorboard-data-server=0.6.0=py38h2b97feb_0 261 | - tensorboard-plugin-wit=1.8.1=pyhd8ed1ab_0 262 | - terminado=0.16.0=pyh41d4057_0 263 | - testpath=0.6.0=pyhd8ed1ab_0 264 | - threadpoolctl=2.2.0=pyh0d69192_0 265 | - tifffile=2021.7.2=pyhd3eb1b0_2 266 | - tk=8.6.11=h1ccaba5_0 267 | - toolz=0.11.2=pyhd3eb1b0_0 268 | - torchaudio=0.11.0=py38_cu113 269 | - torchvision=0.12.0=py38_cu113 270 | - tornado=6.1=py38h0a891b7_3 271 | - tqdm=4.63.0=pyhd3eb1b0_0 272 | - traitlets=5.4.0=pyhd8ed1ab_0 273 | - typing-extensions=4.1.1=hd3eb1b0_0 274 | - typing_extensions=4.1.1=pyh06a4308_0 275 | - urllib3=1.26.8=pyhd3eb1b0_0 276 | - wandb=0.14.2=pyhd8ed1ab_0 277 | - wcwidth=0.2.5=pyh9f0ad1d_2 278 | - webencodings=0.5.1=py_1 279 | - werkzeug=2.1.1=pyhd8ed1ab_0 280 | - wheel=0.37.1=pyhd3eb1b0_0 281 | - widgetsnbextension=4.0.3=pyhd8ed1ab_0 282 | - xorg-kbproto=1.0.7=h7f98852_1002 283 | - xorg-libice=1.0.10=h7f98852_0 284 | - xorg-libsm=1.2.3=hd9c2040_1000 285 | - xorg-libx11=1.7.2=h7f98852_0 286 | - xorg-libxau=1.0.9=h7f98852_0 287 | - xorg-libxdmcp=1.1.3=h7f98852_0 288 | - xorg-libxext=1.3.4=h7f98852_1 289 | - xorg-libxrender=0.9.10=h7f98852_1003 290 | - xorg-renderproto=0.11.1=h7f98852_1002 291 | - xorg-xextproto=7.3.0=h7f98852_1002 292 | - xorg-xproto=7.0.31=h7f98852_1007 293 | - xz=5.2.5=h7b6447c_0 294 | - yaml=0.2.5=h7b6447c_0 295 | - yarl=1.6.3=py38h27cfd23_0 296 | - zeromq=4.3.4=h9c3ff4c_1 297 | - zfp=0.5.5=h295c915_6 298 | - zipp=3.8.0=pyhd8ed1ab_0 299 | - zlib=1.2.13=h5eee18b_0 300 | - zstd=1.4.9=haebb681_0 301 | - pip: 302 | - astunparse==1.6.3 303 | - colour-demosaicing==0.2.1 304 | - colour-science==0.4.1 305 | - cupy-cuda113==10.4.0 306 | - fastrlock==0.8 307 | - flatbuffers==1.12 308 | - gast==0.4.0 309 | - google-pasta==0.2.0 310 | - guided-filter-pytorch==3.7.5 311 | - keras==2.9.0 312 | - keras-preprocessing==1.1.2 313 | - libclang==14.0.1 314 | - mat73==0.60 315 | - opt-einsum==3.3.0 316 | - pip==22.1.1 317 | - ptflops==0.6.9 318 | - pyqt5-sip==4.19.18 319 | - pyqtchart==5.12 320 | - pyqtwebengine==5.12.1 321 | - pyyaml==6.0 322 | - rawpy==0.17.1 323 | - tensorboard==2.9.1 324 | - tensorflow==2.9.1 325 | - tensorflow-estimator==2.9.0 326 | - tensorflow-io-gcs-filesystem==0.26.0 327 | - termcolor==1.1.0 328 | - thop==0.0.31-2005241907 329 | - timm==0.3.2 330 | - wrapt==1.14.1 331 | prefix: /hdd/Documents/anaconda/image_denoising 332 | -------------------------------------------------------------------------------- /model/__init__.py: -------------------------------------------------------------------------------- 1 | from model.bnn import BNNModel 2 | from model.three_stage import ThreeStageModel -------------------------------------------------------------------------------- /model/base.py: -------------------------------------------------------------------------------- 1 | from abc import abstractmethod 2 | import os 3 | import torch 4 | from torch.nn.parallel import DataParallel 5 | from utils.build import build 6 | from utils.io import log 7 | 8 | class BaseModel(): 9 | def __init__(self, opt): 10 | self.opt = opt 11 | self.iter = 0 if 'iter' not in opt else opt['iter'] 12 | self.networks = {} 13 | for network_opt in opt['networks']: 14 | Net = getattr(__import__('network'), network_opt['type']) 15 | net = build(Net, network_opt['args']) 16 | if 'path' in network_opt.keys(): 17 | self.load_net(net, network_opt['path']) 18 | self.networks[network_opt['name']] = net 19 | 20 | @abstractmethod 21 | def train_step(self, data): 22 | pass 23 | 24 | @abstractmethod 25 | def validation_step(self, data): 26 | pass 27 | 28 | def data_parallel(self): 29 | for name in self.networks.keys(): 30 | net = self.networks[name] 31 | net = net.cuda() 32 | net = DataParallel(net) 33 | self.networks[name] = net 34 | 35 | def save_net(self): 36 | for name, net in self.networks.items(): 37 | if isinstance(net, DataParallel): 38 | net = net.module 39 | torch.save(net.state_dict(), os.path.join(self.opt['log_dir'], '%s_iter_%08d.pth' % (name, self.iter))) 40 | 41 | def load_net(self, net, path): 42 | state_dict = torch.load(path) 43 | net.load_state_dict(state_dict) 44 | 45 | @abstractmethod 46 | def save_model(self): 47 | pass 48 | 49 | @abstractmethod 50 | def load_model(self, path): 51 | pass 52 | 53 | def log(self): 54 | log(self.opt['log_file'], 'iter: %d, loss: %f\n' % (self.iter, self.loss.item())) -------------------------------------------------------------------------------- /model/bnn.py: -------------------------------------------------------------------------------- 1 | from model.base import BaseModel 2 | import os 3 | import torch 4 | import torch.nn as nn 5 | 6 | class BNNModel(BaseModel): 7 | def __init__(self, opt): 8 | super(BNNModel, self).__init__(opt) 9 | self.criteron = nn.L1Loss(reduction='mean') 10 | self.optimizer = torch.optim.Adam(self.networks['BNN'].parameters(), lr=opt['lr']) 11 | self.scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(self.optimizer, opt['num_iters']) 12 | 13 | def train_step(self, data): 14 | input = data['L'] 15 | self.networks['BNN'].train() 16 | output = self.networks['BNN'](input) 17 | 18 | self.loss = self.criteron(output, input) 19 | self.optimizer.zero_grad() 20 | self.loss.backward() 21 | self.optimizer.step() 22 | self.scheduler.step() 23 | self.iter += 1 24 | 25 | def validation_step(self, data): 26 | input = data['L'] 27 | self.networks['BNN'].eval() 28 | with torch.no_grad(): 29 | output = self.networks['BNN'](input) 30 | 31 | return output 32 | 33 | def save_model(self): 34 | save_dict = {'iter': self.iter, 35 | 'optimizer': self.optimizer.state_dict(), 36 | 'scheduler': self.scheduler.state_dict(), 37 | 'BNN': self.networks['BNN'].state_dict()} 38 | torch.save(save_dict, os.path.join(self.opt['log_dir'], 'model_iter_%08d.pth' % self.iter)) 39 | 40 | def load_model(self, path): 41 | load_dict = torch.load(path) 42 | self.iter = load_dict['iter'] 43 | self.optimizer.load_state_dict(load_dict['optimizer']) 44 | self.scheduler.load_state_dict(load_dict['scheduler']) 45 | self.networks['BNN'].load_state_dict(load_dict['BNN']) 46 | 47 | 48 | -------------------------------------------------------------------------------- /model/three_stage.py: -------------------------------------------------------------------------------- 1 | from model.base import BaseModel 2 | import os 3 | import torch 4 | import torch.nn as nn 5 | from torch.nn.parallel import DataParallel 6 | 7 | def std(img, window_size=7): 8 | assert window_size % 2 == 1 9 | pad = window_size // 2 10 | 11 | # calculate std on the mean image of the color channels 12 | img = torch.mean(img, dim=1, keepdim=True) 13 | N, C, H, W = img.shape 14 | img = nn.functional.pad(img, [pad] * 4, mode='reflect') 15 | img = nn.functional.unfold(img, kernel_size=window_size) 16 | img = img.view(N, C, window_size * window_size, H, W) 17 | img = img - torch.mean(img, dim=2, keepdim=True) 18 | img = img * img 19 | img = torch.mean(img, dim=2, keepdim=True) 20 | img = torch.sqrt(img) 21 | img = img.squeeze(2) 22 | return img 23 | 24 | def generate_alpha(input, lower=1, upper=5): 25 | N, C, H, W = input.shape 26 | ratio = input.new_ones((N, 1, H, W)) * 0.5 27 | input_std = std(input) 28 | ratio[input_std < lower] = torch.sigmoid((input_std - lower))[input_std < lower] 29 | ratio[input_std > upper] = torch.sigmoid((input_std - upper))[input_std > upper] 30 | ratio = ratio.detach() 31 | 32 | return ratio 33 | 34 | class ThreeStageModel(BaseModel): 35 | def __init__(self, opt): 36 | super(ThreeStageModel, self).__init__(opt) 37 | self.stage = None 38 | self.criteron = nn.L1Loss(reduction='mean') 39 | self.optimizer_BNN = torch.optim.Adam(self.networks['BNN'].parameters(), lr=opt['lr']) 40 | self.scheduler_BNN = torch.optim.lr_scheduler.CosineAnnealingLR(self.optimizer_BNN, opt['BNN_iters']) 41 | self.optimizer_LAN = torch.optim.Adam(self.networks['LAN'].parameters(), lr=opt['lr']) 42 | self.scheduler_LAN = torch.optim.lr_scheduler.CosineAnnealingLR(self.optimizer_LAN, opt['LAN_iters']) 43 | self.optimizer_UNet = torch.optim.Adam(self.networks['UNet'].parameters(), lr=opt['lr']) 44 | self.scheduler_UNet = torch.optim.lr_scheduler.CosineAnnealingLR(self.optimizer_UNet, opt['UNet_iters']) 45 | 46 | def train_step(self, data): 47 | self.iter += 1 48 | self.update_stage() 49 | 50 | input = data['L'] 51 | 52 | if self.stage == 'BNN': 53 | self.networks['BNN'].train() 54 | BNN = self.networks['BNN'](input) 55 | self.loss = self.criteron(BNN, input) 56 | self.optimizer_BNN.zero_grad() 57 | self.loss.backward() 58 | self.optimizer_BNN.step() 59 | self.scheduler_BNN.step() 60 | 61 | elif self.stage == 'LAN': 62 | self.networks['BNN'].eval() 63 | self.networks['LAN'].train() 64 | with torch.no_grad(): 65 | BNN = self.networks['BNN'](input) 66 | LAN = self.networks['LAN'](input) 67 | 68 | # alpha = generate_alpha(BNN) 69 | # self.loss = self.criteron(BNN.detach() * (1 - alpha), LAN * (1 - alpha)) 70 | self.loss = self.criteron(BNN, LAN) 71 | self.optimizer_LAN.zero_grad() 72 | self.loss.backward() 73 | self.optimizer_LAN.step() 74 | self.scheduler_LAN.step() 75 | 76 | elif self.stage == 'UNet': 77 | self.networks['BNN'].eval() 78 | self.networks['LAN'].eval() 79 | self.networks['UNet'].train() 80 | with torch.no_grad(): 81 | BNN = self.networks['BNN'](input) 82 | LAN = self.networks['LAN'](input) 83 | UNet = self.networks['UNet'](input) 84 | 85 | alpha = generate_alpha(BNN) 86 | self.loss = self.criteron(BNN * (1 - alpha), UNet * (1 - alpha)) + self.criteron(LAN * alpha, UNet * alpha) 87 | self.optimizer_UNet.zero_grad() 88 | self.loss.backward() 89 | self.optimizer_UNet.step() 90 | self.scheduler_UNet.step() 91 | 92 | 93 | def validation_step(self, data): 94 | self.update_stage() 95 | input = data['L'] 96 | 97 | if self.stage == 'BNN': 98 | self.networks['BNN'].eval() 99 | with torch.no_grad(): 100 | output = self.networks['BNN'](input) 101 | elif self.stage == 'LAN': 102 | self.networks['LAN'].eval() 103 | with torch.no_grad(): 104 | output = self.networks['LAN'](input) 105 | elif self.stage == 'UNet': 106 | self.networks['UNet'].eval() 107 | with torch.no_grad(): 108 | output = self.networks['UNet'](input) 109 | 110 | return output 111 | 112 | def save_net(self): 113 | if self.stage == 'BNN': 114 | net = self.networks['BNN'] 115 | elif self.stage == 'LAN': 116 | net = self.networks['LAN'] 117 | elif self.stage == 'UNet': 118 | net = self.networks['UNet'] 119 | 120 | if isinstance(net, DataParallel): 121 | net = net.module 122 | torch.save(net.state_dict(), os.path.join(self.opt['log_dir'], 'net_iter_%08d.pth' % self.iter)) 123 | 124 | def save_model(self): 125 | if self.stage == 'BNN': 126 | save_dict = {'iter': self.iter, 127 | 'optimizer_BNN': self.optimizer_BNN.state_dict(), 128 | 'scheduler_BNN': self.scheduler_BNN.state_dict(), 129 | 'BNN': self.networks['BNN'].state_dict()} 130 | elif self.stage == 'LAN': 131 | save_dict = {'iter': self.iter, 132 | 'optimizer_LAN': self.optimizer_LAN.state_dict(), 133 | 'scheduler_LAN': self.scheduler_LAN.state_dict(), 134 | 'BNN': self.networks['BNN'].state_dict(), 135 | 'LAN': self.networks['LAN'].state_dict()} 136 | elif self.stage == 'UNet': 137 | save_dict = {'iter': self.iter, 138 | 'optimizer_UNet': self.optimizer_UNet.state_dict(), 139 | 'scheduler_UNet': self.scheduler_UNet.state_dict(), 140 | 'BNN': self.networks['BNN'].state_dict(), 141 | 'LAN': self.networks['LAN'].state_dict(), 142 | 'UNet': self.networks['UNet'].state_dict()} 143 | torch.save(save_dict, os.path.join(self.opt['log_dir'], 'model_iter_%08d.pth' % self.iter)) 144 | 145 | def load_model(self, path): 146 | load_dict = torch.load(path) 147 | self.iter = load_dict['iter'] 148 | self.update_stage() 149 | if self.stage == 'BNN': 150 | self.optimizer_BNN.load_state_dict(load_dict['optimizer_BNN']) 151 | self.scheduler_BNN.load_state_dict(load_dict['scheduler_BNN']) 152 | self.networks['BNN'].load_state_dict(load_dict['BNN']) 153 | elif self.stage == 'LAN': 154 | self.optimizer_LAN.load_state_dict(load_dict['optimizer_LAN']) 155 | self.scheduler_LAN.load_state_dict(load_dict['scheduler_LAN']) 156 | self.networks['BNN'].load_state_dict(load_dict['BNN']) 157 | self.networks['LAN'].load_state_dict(load_dict['LAN']) 158 | elif self.stage == 'UNet': 159 | self.optimizer_UNet.load_state_dict(load_dict['optimizer_UNet']) 160 | self.scheduler_UNet.load_state_dict(load_dict['scheduler_UNet']) 161 | self.networks['BNN'].load_state_dict(load_dict['BNN']) 162 | self.networks['LAN'].load_state_dict(load_dict['LAN']) 163 | self.networks['UNet'].load_state_dict(load_dict['UNet']) 164 | else: 165 | raise NotImplementedError 166 | 167 | def update_stage(self): 168 | if self.iter <= self.opt['BNN_iters']: 169 | self.stage = 'BNN' 170 | elif self.iter <= self.opt['BNN_iters'] + self.opt['LAN_iters']: 171 | self.stage = 'LAN' 172 | else: 173 | self.stage = 'UNet' 174 | -------------------------------------------------------------------------------- /network/__init__.py: -------------------------------------------------------------------------------- 1 | from network.bnn import BNN 2 | from network.lan import LAN 3 | from network.unet import UNet -------------------------------------------------------------------------------- /network/bnn.py: -------------------------------------------------------------------------------- 1 | ''' 2 | The code of BNN is modified from https://github.com/COMP6248-Reproducability-Challenge/selfsupervised-denoising/blob/master-with-report/ssdn/ssdn/models/noise_network.py 3 | ''' 4 | import torch 5 | import torch.nn as nn 6 | from typing import Tuple 7 | 8 | def rotate(x, angle): 9 | """Rotate images by 90 degrees clockwise. Can handle any 2D data format. 10 | Args: 11 | x (Tensor): Image or batch of images. 12 | angle (int): Clockwise rotation angle in multiples of 90. 13 | data_format (str, optional): Format of input image data, e.g. BCHW, 14 | HWC. Defaults to BCHW. 15 | Returns: 16 | Tensor: Copy of tensor with rotation applied. 17 | """ 18 | h_dim, w_dim = 2, 3 19 | 20 | if angle == 0: 21 | return x 22 | elif angle == 90: 23 | return x.flip(w_dim).transpose(h_dim, w_dim) 24 | elif angle == 180: 25 | return x.flip(w_dim).flip(h_dim) 26 | elif angle == 270: 27 | return x.flip(h_dim).transpose(h_dim, w_dim) 28 | else: 29 | raise NotImplementedError("Must be rotation divisible by 90 degrees") 30 | 31 | class Crop2d(nn.Module): 32 | """Crop input using slicing. Assumes BCHW data. 33 | 34 | Args: 35 | crop (Tuple[int, int, int, int]): Amounts to crop from each side of the image. 36 | Tuple is treated as [left, right, top, bottom]/ 37 | """ 38 | 39 | def __init__(self, crop: Tuple[int, int, int, int]): 40 | super().__init__() 41 | self.crop = crop 42 | assert len(crop) == 4 43 | 44 | def forward(self, x): 45 | (left, right, top, bottom) = self.crop 46 | x0, x1 = left, x.shape[-1] - right 47 | y0, y1 = top, x.shape[-2] - bottom 48 | return x[:, :, y0:y1, x0:x1] 49 | 50 | 51 | class Shift2d(nn.Module): 52 | """Shift an image in either or both of the vertical and horizontal axis by first 53 | zero padding on the opposite side that the image is shifting towards before 54 | cropping the side being shifted towards. 55 | 56 | Args: 57 | shift (Tuple[int, int]): Tuple of vertical and horizontal shift. Positive values 58 | shift towards right and bottom, negative values shift towards left and top. 59 | """ 60 | 61 | def __init__(self, shift: Tuple[int, int]): 62 | super().__init__() 63 | self.shift = shift 64 | vert, horz = self.shift 65 | y_a, y_b = abs(vert), 0 66 | x_a, x_b = abs(horz), 0 67 | if vert < 0: 68 | y_a, y_b = y_b, y_a 69 | if horz < 0: 70 | x_a, x_b = x_b, x_a 71 | # Order : Left, Right, Top Bottom 72 | self.pad = nn.ZeroPad2d((x_a, x_b, y_a, y_b)) 73 | self.crop = Crop2d((x_b, x_a, y_b, y_a)) 74 | self.shift_block = nn.Sequential(self.pad, self.crop) 75 | 76 | def forward(self, x): 77 | return self.shift_block(x) 78 | 79 | class ShiftConv2d(nn.Conv2d): 80 | def __init__(self, *args, **kwargs): 81 | """Custom convolution layer as defined by Laine et al. for restricting the 82 | receptive field of a convolution layer to only be upwards. For a h × w kernel, 83 | a downwards offset of k = [h/2] pixels is used. This is applied as a k sized pad 84 | to the top of the input before applying the convolution. The bottom k rows are 85 | cropped out for output. 86 | """ 87 | super().__init__(*args, **kwargs) 88 | self.shift_size = (self.kernel_size[0] // 2, 0) 89 | # Use individual layers of shift for wrapping conv with shift 90 | shift = Shift2d(self.shift_size) 91 | self.pad = shift.pad 92 | self.crop = shift.crop 93 | 94 | def forward(self, x): 95 | x = self.pad(x) 96 | x = super().forward(x) 97 | x = self.crop(x) 98 | return x 99 | 100 | 101 | class BNN(nn.Module): 102 | def __init__(self, blindspot, in_ch=3, out_ch=3, dim=48): 103 | super(BNN, self).__init__() 104 | in_channels = in_ch 105 | out_channels = out_ch 106 | self.blindspot = blindspot 107 | 108 | #################################### 109 | # Encode Blocks 110 | #################################### 111 | 112 | # Layers: enc_conv0, enc_conv1, pool1 113 | self.encode_block_1 = nn.Sequential( 114 | ShiftConv2d(in_channels, dim, 3, stride=1, padding=1), 115 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 116 | ShiftConv2d(dim, dim, 3, padding=1), 117 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 118 | Shift2d((1, 0)), 119 | nn.MaxPool2d(2) 120 | ) 121 | 122 | # Layers: enc_conv(i), pool(i); i=2..5 123 | def _encode_block_2_3_4_5() -> nn.Module: 124 | return nn.Sequential( 125 | ShiftConv2d(dim, dim, 3, stride=1, padding=1), 126 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 127 | Shift2d((1, 0)), 128 | nn.MaxPool2d(2) 129 | ) 130 | 131 | # Separate instances of same encode module definition created 132 | self.encode_block_2 = _encode_block_2_3_4_5() 133 | self.encode_block_3 = _encode_block_2_3_4_5() 134 | self.encode_block_4 = _encode_block_2_3_4_5() 135 | self.encode_block_5 = _encode_block_2_3_4_5() 136 | 137 | # Layers: enc_conv6 138 | self.encode_block_6 = nn.Sequential( 139 | ShiftConv2d(dim, dim, 3, stride=1, padding=1), 140 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 141 | ) 142 | 143 | #################################### 144 | # Decode Blocks 145 | #################################### 146 | # Layers: upsample5 147 | self.decode_block_6 = nn.Sequential(nn.Upsample(scale_factor=2, mode="nearest")) 148 | 149 | # Layers: dec_conv5a, dec_conv5b, upsample4 150 | self.decode_block_5 = nn.Sequential( 151 | ShiftConv2d(2 * dim, 2 * dim, 3, stride=1, padding=1), 152 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 153 | ShiftConv2d(2 * dim, 2 * dim, 3, stride=1, padding=1), 154 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 155 | nn.Upsample(scale_factor=2, mode="nearest"), 156 | ) 157 | 158 | # Layers: dec_deconv(i)a, dec_deconv(i)b, upsample(i-1); i=4..2 159 | def _decode_block_4_3_2() -> nn.Module: 160 | return nn.Sequential( 161 | ShiftConv2d(3 * dim, 2 * dim, 3, stride=1, padding=1), 162 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 163 | ShiftConv2d(2 * dim, 2 * dim, 3, stride=1, padding=1), 164 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 165 | nn.Upsample(scale_factor=2, mode="nearest"), 166 | ) 167 | 168 | # Separate instances of same decode module definition created 169 | self.decode_block_4 = _decode_block_4_3_2() 170 | self.decode_block_3 = _decode_block_4_3_2() 171 | self.decode_block_2 = _decode_block_4_3_2() 172 | 173 | # Layers: dec_conv1a, dec_conv1b, dec_conv1c, 174 | self.decode_block_1 = nn.Sequential( 175 | ShiftConv2d(2 * dim + in_channels, 2 * dim, 3, stride=1, padding=1), 176 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 177 | ShiftConv2d(2 * dim, 2 * dim, 3, stride=1, padding=1), 178 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 179 | ) 180 | 181 | #################################### 182 | # Output Block 183 | #################################### 184 | 185 | # Shift blindspot pixel down 186 | self.shift = Shift2d(((self.blindspot + 1) // 2, 0)) 187 | 188 | # nin_a,b,c, linear_act 189 | self.output_conv = ShiftConv2d(2 * dim, out_channels, 1) 190 | self.output_block = nn.Sequential( 191 | ShiftConv2d(8 * dim, 8 * dim, 1), 192 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 193 | ShiftConv2d(8 * dim, 2 * dim, 1), 194 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 195 | self.output_conv, 196 | ) 197 | 198 | # Initialize weights 199 | self.init_weights() 200 | 201 | def init_weights(self): 202 | """Initializes weights using Kaiming He et al. (2015). 203 | 204 | Only convolution layers have learnable weights. All convolutions use a leaky 205 | relu activation function (negative_slope = 0.1) except the last which is just 206 | a linear output. 207 | """ 208 | with torch.no_grad(): 209 | self._init_weights() 210 | 211 | def _init_weights(self): 212 | for m in self.modules(): 213 | if isinstance(m, nn.Conv2d): 214 | nn.init.kaiming_normal_(m.weight.data, a=0.1) 215 | m.bias.data.zero_() 216 | # Initialise last output layer 217 | nn.init.kaiming_normal_(self.output_conv.weight.data, nonlinearity="linear") 218 | 219 | def forward(self, x, shift=None): 220 | if shift is not None: 221 | self.shift = Shift2d((shift, 0)) 222 | else: 223 | self.shift = Shift2d(((self.blindspot + 1) // 2, 0)) 224 | 225 | rotated = [rotate(x, rot) for rot in (0, 90, 180, 270)] 226 | x = torch.cat((rotated), dim=0) 227 | 228 | # Encoder 229 | pool1 = self.encode_block_1(x) 230 | pool2 = self.encode_block_2(pool1) 231 | pool3 = self.encode_block_3(pool2) 232 | pool4 = self.encode_block_4(pool3) 233 | pool5 = self.encode_block_5(pool4) 234 | encoded = self.encode_block_6(pool5) 235 | 236 | # Decoder 237 | upsample5 = self.decode_block_6(encoded) 238 | concat5 = torch.cat((upsample5, pool4), dim=1) 239 | upsample4 = self.decode_block_5(concat5) 240 | concat4 = torch.cat((upsample4, pool3), dim=1) 241 | upsample3 = self.decode_block_4(concat4) 242 | concat3 = torch.cat((upsample3, pool2), dim=1) 243 | upsample2 = self.decode_block_3(concat3) 244 | concat2 = torch.cat((upsample2, pool1), dim=1) 245 | upsample1 = self.decode_block_2(concat2) 246 | concat1 = torch.cat((upsample1, x), dim=1) 247 | x = self.decode_block_1(concat1) 248 | 249 | # Apply shift 250 | shifted = self.shift(x) 251 | # Unstack, rotate and combine 252 | rotated_batch = torch.chunk(shifted, 4, dim=0) 253 | aligned = [ 254 | rotate(rotated, rot) 255 | for rotated, rot in zip(rotated_batch, (0, 270, 180, 90)) 256 | ] 257 | x = torch.cat(aligned, dim=1) 258 | 259 | x = self.output_block(x) 260 | 261 | return x 262 | 263 | @staticmethod 264 | def input_wh_mul() -> int: 265 | """Multiple that both the width and height dimensions of an input must be to be 266 | processed by the network. This is devised from the number of pooling layers that 267 | reduce the input size. 268 | 269 | Returns: 270 | int: Dimension multiplier 271 | """ 272 | max_pool_layers = 5 273 | return 2 ** max_pool_layers -------------------------------------------------------------------------------- /network/lan.py: -------------------------------------------------------------------------------- 1 | import torch 2 | import torch.nn as nn 3 | 4 | class CALayer(nn.Module): 5 | def __init__(self, channel=64, reduction=16): 6 | super(CALayer, self).__init__() 7 | 8 | self.avg_pool = nn.AdaptiveAvgPool2d(1) 9 | self.conv_du = nn.Sequential( 10 | nn.Conv2d(channel, channel // reduction, 1, padding=0, bias=True), 11 | nn.ReLU(inplace=True), 12 | nn.Conv2d(channel // reduction, channel, 1, padding=0, bias=True), 13 | nn.Sigmoid() 14 | ) 15 | 16 | def forward(self, x): 17 | y = self.avg_pool(x) 18 | y = self.conv_du(y) 19 | return x * y 20 | 21 | class RB(nn.Module): 22 | def __init__(self, filters): 23 | super(RB, self).__init__() 24 | self.conv1 = nn.Conv2d(filters, filters, 1) 25 | self.act = nn.ReLU() 26 | self.conv2 = nn.Conv2d(filters, filters, 1) 27 | self.cuca = CALayer(channel=filters) 28 | 29 | def forward(self, x): 30 | c0 = x 31 | x = self.conv1(x) 32 | x = self.act(x) 33 | x = self.conv2(x) 34 | out = self.cuca(x) 35 | return out + c0 36 | 37 | class NRB(nn.Module): 38 | def __init__(self, n, filters): 39 | super(NRB, self).__init__() 40 | nets = [] 41 | for i in range(n): 42 | nets.append(RB(filters)) 43 | self.body = nn.Sequential(*nets) 44 | self.tail = nn.Conv2d(filters, filters, 1) 45 | 46 | def forward(self, x): 47 | return x + self.tail(self.body(x)) 48 | 49 | 50 | class LAN(nn.Module): 51 | def __init__(self, blindspot, in_ch=3, out_ch=None, rbs=6): 52 | super(LAN, self).__init__() 53 | self.receptive_feild = blindspot 54 | assert self.receptive_feild % 2 == 1 55 | self.in_ch = in_ch 56 | self.out_ch = self.in_ch if out_ch is None else out_ch 57 | self.mid_ch = 64 58 | self.rbs = rbs 59 | 60 | layers = [] 61 | layers.append(nn.Conv2d(self.in_ch, self.mid_ch, 1)) 62 | layers.append(nn.ReLU()) 63 | 64 | for i in range(self.receptive_feild // 2): 65 | layers.append(nn.Conv2d(self.mid_ch, self.mid_ch, 3, 1, 1)) 66 | layers.append(nn.ReLU()) 67 | 68 | layers.append(NRB(self.rbs, self.mid_ch)) 69 | layers.append(nn.Conv2d(self.mid_ch, self.out_ch, 1)) 70 | 71 | self.conv = nn.Sequential(*layers) 72 | 73 | def forward(self, x): 74 | return self.conv(x) 75 | -------------------------------------------------------------------------------- /network/unet.py: -------------------------------------------------------------------------------- 1 | ''' 2 | U-Net is also modified from https://github.com/COMP6248-Reproducability-Challenge/selfsupervised-denoising/blob/master-with-report/ssdn/ssdn/models/noise_network.py 3 | ''' 4 | import torch 5 | import torch.nn as nn 6 | from typing import Tuple 7 | 8 | class UNet(nn.Module): 9 | def __init__(self, in_ch=3, out_ch=3, zero_output=False, dim=48): 10 | super(UNet, self).__init__() 11 | self.zero_output = zero_output 12 | in_channels = in_ch 13 | out_channels = out_ch 14 | 15 | #################################### 16 | # Encode Blocks 17 | #################################### 18 | 19 | # Layers: enc_conv0, enc_conv1, pool1 20 | self.encode_block_1 = nn.Sequential( 21 | nn.Conv2d(in_channels, dim, 3, stride=1, padding=1), 22 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 23 | nn.Conv2d(dim, dim, 3, padding=1), 24 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 25 | nn.MaxPool2d(2) 26 | ) 27 | 28 | # Layers: enc_conv(i), pool(i); i=2..5 29 | def _encode_block_2_3_4_5() -> nn.Module: 30 | return nn.Sequential( 31 | nn.Conv2d(dim, dim, 3, stride=1, padding=1), 32 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 33 | nn.MaxPool2d(2) 34 | ) 35 | 36 | # Separate instances of same encode module definition created 37 | self.encode_block_2 = _encode_block_2_3_4_5() 38 | self.encode_block_3 = _encode_block_2_3_4_5() 39 | self.encode_block_4 = _encode_block_2_3_4_5() 40 | self.encode_block_5 = _encode_block_2_3_4_5() 41 | 42 | # Layers: enc_conv6 43 | self.encode_block_6 = nn.Sequential( 44 | nn.Conv2d(dim, dim, 3, stride=1, padding=1), 45 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 46 | ) 47 | 48 | #################################### 49 | # Decode Blocks 50 | #################################### 51 | # Layers: upsample5 52 | self.decode_block_6 = nn.Sequential(nn.Upsample(scale_factor=2, mode="nearest")) 53 | 54 | # Layers: dec_conv5a, dec_conv5b, upsample4 55 | self.decode_block_5 = nn.Sequential( 56 | nn.Conv2d(dim * 2, dim * 2, 3, stride=1, padding=1), 57 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 58 | nn.Conv2d(dim * 2, dim * 2, 3, stride=1, padding=1), 59 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 60 | nn.Upsample(scale_factor=2, mode="nearest"), 61 | ) 62 | 63 | # Layers: dec_deconv(i)a, dec_deconv(i)b, upsample(i-1); i=4..2 64 | def _decode_block_4_3_2() -> nn.Module: 65 | return nn.Sequential( 66 | nn.Conv2d(dim * 3, dim * 2, 3, stride=1, padding=1), 67 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 68 | nn.Conv2d(dim * 2, dim * 2, 3, stride=1, padding=1), 69 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 70 | nn.Upsample(scale_factor=2, mode="nearest"), 71 | ) 72 | 73 | # Separate instances of same decode module definition created 74 | self.decode_block_4 = _decode_block_4_3_2() 75 | self.decode_block_3 = _decode_block_4_3_2() 76 | self.decode_block_2 = _decode_block_4_3_2() 77 | 78 | # Layers: dec_conv1a, dec_conv1b, dec_conv1c, 79 | self.decode_block_1 = nn.Sequential( 80 | nn.Conv2d(dim * 2 + in_channels, dim * 2, 3, stride=1, padding=1), 81 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 82 | nn.Conv2d(dim * 2, dim * 2, 3, stride=1, padding=1), 83 | nn.LeakyReLU(negative_slope=0.1, inplace=True), 84 | ) 85 | 86 | #################################### 87 | # Output Block 88 | #################################### 89 | 90 | 91 | # nin_a,b,c, linear_act 92 | self.output_conv = nn.Conv2d(dim * 2, out_channels, 1) 93 | 94 | # Initialize weights 95 | self.init_weights() 96 | 97 | def init_weights(self): 98 | """Initializes weights using Kaiming He et al. (2015). 99 | 100 | Only convolution layers have learnable weights. All convolutions use a leaky 101 | relu activation function (negative_slope = 0.1) except the last which is just 102 | a linear output. 103 | """ 104 | with torch.no_grad(): 105 | self._init_weights() 106 | 107 | def _init_weights(self): 108 | for m in self.modules(): 109 | if isinstance(m, nn.Conv2d): 110 | nn.init.kaiming_normal_(m.weight.data, a=0.1) 111 | m.bias.data.zero_() 112 | 113 | # Initialise last output layer 114 | if self.zero_output: 115 | self.output_conv.weight.zero_() 116 | else: 117 | nn.init.kaiming_normal_(self.output_conv.weight.data, nonlinearity="linear") 118 | 119 | def forward(self, x): 120 | 121 | # Encoder 122 | pool1 = self.encode_block_1(x) 123 | pool2 = self.encode_block_2(pool1) 124 | pool3 = self.encode_block_3(pool2) 125 | pool4 = self.encode_block_4(pool3) 126 | pool5 = self.encode_block_5(pool4) 127 | encoded = self.encode_block_6(pool5) 128 | 129 | # Decoder 130 | upsample5 = self.decode_block_6(encoded) 131 | concat5 = torch.cat((upsample5, pool4), dim=1) 132 | upsample4 = self.decode_block_5(concat5) 133 | concat4 = torch.cat((upsample4, pool3), dim=1) 134 | upsample3 = self.decode_block_4(concat4) 135 | concat3 = torch.cat((upsample3, pool2), dim=1) 136 | upsample2 = self.decode_block_3(concat3) 137 | concat2 = torch.cat((upsample2, pool1), dim=1) 138 | upsample1 = self.decode_block_2(concat2) 139 | concat1 = torch.cat((upsample1, x), dim=1) 140 | x = self.decode_block_1(concat1) 141 | 142 | x = self.output_conv(x) 143 | 144 | return x 145 | 146 | @staticmethod 147 | def input_wh_mul() -> int: 148 | """Multiple that both the width and height dimensions of an input must be to be 149 | processed by the network. This is devised from the number of pooling layers that 150 | reduce the input size. 151 | 152 | Returns: 153 | int: Dimension multiplier 154 | """ 155 | max_pool_layers = 5 156 | return 2 ** max_pool_layers -------------------------------------------------------------------------------- /option/bnn.json: -------------------------------------------------------------------------------- 1 | { 2 | // model 3 | "model": "BNNModel", 4 | // net 5 | "networks": [{ 6 | "name": "BNN", 7 | "type": "BNN", 8 | "args": { 9 | "blindspot": 9 10 | } 11 | ,"path": "/home/nagejacob/Documents/codes/SpatiallyAdaptiveSSID/pretrained_models/BNN.pth" 12 | }], 13 | // datasets 14 | "train_datasets": [{ 15 | "type": "SIDDSrgbTrainDataset", 16 | "args": { 17 | "patch_size": 256, 18 | "pin_memory": true 19 | }, 20 | "batch_size": 8 21 | }], 22 | "validation_datasets": [{ 23 | "type": "SIDDSrgbValidationDataset", 24 | "args": {} 25 | }], 26 | // training parameters 27 | "lr": 3e-4, 28 | "print_every": 1000000, 29 | "save_every": 10000, 30 | "validate_every": 10000, 31 | "num_iters": 400000, 32 | "log_dir": "logs", 33 | "log_file": "logs/log.out" 34 | // , "resume_from": "" 35 | } 36 | -------------------------------------------------------------------------------- /option/three_stage.json: -------------------------------------------------------------------------------- 1 | { 2 | // model 3 | "model": "ThreeStageModel", 4 | "iter": 1200000, 5 | // net 6 | "networks":[{ 7 | "name": "BNN", 8 | "type": "BNN", 9 | "args": { 10 | "blindspot": 9 11 | } 12 | // , "path": "/home/nagejacob/Documents/codes/SpatiallyAdaptiveSSID/pretrained_models/BNN.pth" 13 | }, { 14 | "name": "LAN", 15 | "type": "LAN", 16 | "args": { 17 | "blindspot": 3 18 | } 19 | // , "path": "/home/nagejacob/Documents/codes/SpatiallyAdaptiveSSID/pretrained_models/LAN.pth" 20 | }, { 21 | "name": "UNet", 22 | "type": "UNet", 23 | "args": {} 24 | , "path": "/home/nagejacob/Documents/codes/SpatiallyAdaptiveSSID/pretrained_models/UNet.pth" 25 | }], 26 | // datasets 27 | "train_datasets": [{ 28 | "type": "SIDDSrgbTrainDataset", 29 | "args": { 30 | "patch_size": 256, 31 | "pin_memory": true 32 | }, 33 | "batch_size": 8 34 | }], 35 | "validation_datasets": [{ 36 | "type": "SIDDSrgbValidationDataset", 37 | "args": {} 38 | }], 39 | // training parameters 40 | "lr": 3e-4, 41 | "print_every": 10000000, 42 | "save_every": 10000, 43 | "validate_every": 10000, 44 | "BNN_iters": 400000, 45 | "LAN_iters": 400000, 46 | "UNet_iters": 400000, 47 | "num_iters": 1200000, 48 | "log_dir": "logs", 49 | "log_file": "logs/log.out" 50 | // , "resume_from": "" 51 | } 52 | -------------------------------------------------------------------------------- /pretrained_models/BNN.pth: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/nagejacob/SpatiallyAdaptiveSSID/7e94712f2b38fd926b0b6e879e9d85f09e942cc4/pretrained_models/BNN.pth -------------------------------------------------------------------------------- /pretrained_models/LAN.pth: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/nagejacob/SpatiallyAdaptiveSSID/7e94712f2b38fd926b0b6e879e9d85f09e942cc4/pretrained_models/LAN.pth -------------------------------------------------------------------------------- /pretrained_models/UNet.pth: -------------------------------------------------------------------------------- https://raw.githubusercontent.com/nagejacob/SpatiallyAdaptiveSSID/7e94712f2b38fd926b0b6e879e9d85f09e942cc4/pretrained_models/UNet.pth -------------------------------------------------------------------------------- /readme.md: -------------------------------------------------------------------------------- 1 | # Spatially Adaptive Self-Supervised Learning for Real-World Image Denoising 2 | The source code for paper "[Spatially Adaptive Self-Supervised Learning for Real-World Image Denoising](https://arxiv.org/pdf/2303.14934.pdf)" (CVPR 2023) 3 | 4 | ## Usage 5 | ### Datasets 6 | Download [SIDD](https://www.eecs.yorku.ca/~kamel/sidd/dataset.php) and [DND](https://noise.visinf.tu-darmstadt.de/) datasets, and modify `dataset_path` in `dataset/base.py` accordingly. 7 | ``` 8 | |- dataset_path 9 | |- SIDD 10 | |- SIDD_Medium_Srgb 11 | |- Data 12 | |- 0001_001_S6_00100_00060_3200_L 13 | |- 0002_001_S6_00100_00020_3200_N 14 | |- ... 15 | |- SIDD_Validation 16 | |- ValidationNoisyBlocksSrgb.mat 17 | |- ValidationGtBlocksSrgb.mat 18 | |- SIDD_Benchmark 19 | |- BenchmarkNoisyBlocksSrgb.mat 20 | |- DND 21 | |- info.mat 22 | |- images_srgb 23 | ``` 24 | 25 | ### Validation 26 | Validate on SIDD Validation dataset, 27 | ``` 28 | cd validate 29 | python validate_SIDD.py 30 | ``` 31 | 32 | ### Training (removed due to confidentiality agreement, see [here](https://github.com/nagejacob/SpatiallyAdaptiveSSID/tree/731adb9b5dcc3a860d207436f5fee6f794b2e5f4)) 33 | Training on SIDD Medium dataset, 34 | ``` 35 | sh train.sh 36 | ``` 37 | 38 | ## Citation 39 | If you find our work useful in your research or publication, please cite: 40 | ``` 41 | @inproceedings{li2023spatially, 42 | title={Spatially Adaptive Self-Supervised Learning for Real-World Image Denoising}, 43 | author={Li, Junyi and Zhang, Zhilu and Liu, Xiaoyu and Feng, Chaoyu and Wang, Xiaotao and Lei, Lei and Zuo, Wangmeng}, 44 | booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition}, 45 | year={2023} 46 | } 47 | ``` 48 | -------------------------------------------------------------------------------- /submit_test/ensemble_wrapper.py: -------------------------------------------------------------------------------- 1 | import torch 2 | 3 | class EnsembleWrapper(): 4 | def __init__(self, model): 5 | self.model = model 6 | 7 | def validation_step(self, data): 8 | input = data['L'] 9 | outputs = [] 10 | for i in range(8): 11 | aug_input = input.clone() 12 | if i >= 4: 13 | aug_input = torch.flip(aug_input, [2]) 14 | if i % 4 > 1: 15 | aug_input = torch.flip(aug_input, [3]) 16 | if (i % 4) % 2 == 1: 17 | aug_input = torch.rot90(aug_input, 1, [2, 3]) 18 | 19 | aug_output = self.model.validation_step({'L': aug_input}) 20 | 21 | if (i % 4) % 2 == 1: 22 | aug_output = torch.rot90(aug_output, 3, [2, 3]) 23 | if i % 4 > 1: 24 | aug_output = torch.flip(aug_output, [3]) 25 | if i >= 4: 26 | aug_output = torch.flip(aug_output, [2]) 27 | outputs.append(aug_output) 28 | output = torch.stack(outputs, dim=0) 29 | output = torch.mean(output, dim=0, keepdim=False) 30 | return output -------------------------------------------------------------------------------- /submit_test/test_DND.py: -------------------------------------------------------------------------------- 1 | import sys 2 | sys.path.append('..') 3 | import argparse 4 | from dataset.DND_sRGB import dnd_path 5 | import h5py 6 | import numpy as np 7 | import os 8 | import scipy.io as sio 9 | import shutil 10 | from submit_test.ensemble_wrapper import EnsembleWrapper 11 | import torch 12 | from tqdm import tqdm 13 | from utils.option import parse, recursive_print 14 | 15 | def bundle_submissions_srgb(submission_folder): 16 | ''' 17 | Bundles submission data for sRGB denoising 18 | 19 | submission_folder Folder where denoised images reside 20 | 21 | Output is written to /bundled/. Please submit 22 | the content of this folder. 23 | ''' 24 | out_folder = os.path.join(submission_folder, "bundled/") 25 | try: 26 | os.mkdir(out_folder) 27 | except: 28 | pass 29 | israw = False 30 | eval_version = "1.0" 31 | 32 | for i in range(50): 33 | Idenoised = np.zeros((20,), dtype=np.object) 34 | for bb in range(20): 35 | filename = '%04d_%02d.mat' % (i + 1, bb + 1) 36 | s = sio.loadmat(os.path.join(submission_folder, filename)) 37 | Idenoised_crop = s["Idenoised_crop"] 38 | Idenoised[bb] = Idenoised_crop 39 | filename = '%04d.mat' % (i + 1) 40 | sio.savemat(os.path.join(out_folder, filename), 41 | {"Idenoised": Idenoised, 42 | "israw": israw, 43 | "eval_version": eval_version}, 44 | ) 45 | 46 | max_margin = 80 47 | 48 | def main(opt): 49 | if opt['save_mat']: 50 | if os.path.exists(opt['mat_dir']): 51 | shutil.rmtree(opt['mat_dir']) 52 | os.makedirs(opt['mat_dir']) 53 | 54 | Model = getattr(__import__('model'), opt['model']) 55 | model = Model(opt) 56 | model.data_parallel() 57 | if 'resume_from' in opt: 58 | model.load_model(opt['resume_from']) 59 | if opt['ensemble']: 60 | model = EnsembleWrapper(model) 61 | 62 | infos = h5py.File(os.path.join(dnd_path, 'info.mat'), 'r') 63 | info = infos['info'] 64 | bb = info['boundingboxes'] 65 | for i in tqdm(range(50)): 66 | filename = os.path.join(dnd_path, 'images_srgb', '%04d.mat' % (i + 1)) 67 | img = h5py.File(filename, 'r') 68 | Inoisy = np.float32(np.array(img['InoisySRGB']).T) 69 | # bounding box 70 | ref = bb[0][i] 71 | boxes = np.array(info[ref]).T 72 | for k in range(20): 73 | idx = [int(boxes[k, 0] - 1), int(boxes[k, 2]), int(boxes[k, 1] - 1), int(boxes[k, 3])] 74 | 75 | # Crop margin for better boundary process 76 | h_min_margin = max_margin 77 | h_max_margin = max_margin 78 | w_min_margin = max_margin 79 | w_max_margin = max_margin 80 | 81 | if 0 > idx[0] - max_margin: 82 | h_min_margin = idx[0] 83 | if Inoisy.shape[0] < idx[1] + max_margin: 84 | h_max_margin = Inoisy.shape[0] - idx[1] 85 | if 0 > idx[2] - max_margin: 86 | w_min_margin = idx[2] 87 | if Inoisy.shape[1] < idx[3] + max_margin: 88 | w_max_margin = Inoisy.shape[1] - idx[3] 89 | 90 | h_min_margin = h_min_margin // 32 * 32 91 | h_max_margin = h_max_margin // 32 * 32 92 | w_min_margin = w_min_margin // 32 * 32 93 | w_max_margin = w_max_margin // 32 * 32 94 | 95 | Inoisy_crop = Inoisy[idx[0] - h_min_margin:idx[1] + h_max_margin, 96 | idx[2] - w_min_margin:idx[3] + w_max_margin, :].copy() 97 | H = Inoisy_crop.shape[0] 98 | W = Inoisy_crop.shape[1] 99 | 100 | Inoisy_crop = torch.from_numpy(Inoisy_crop).permute(2, 0, 1).unsqueeze(0).cuda() 101 | Inoisy_crop = Inoisy_crop * 255. 102 | 103 | Idenoised_crop = model.validation_step({'L': Inoisy_crop}) 104 | 105 | Idenoised_crop = torch.clamp(Idenoised_crop / 255., 0., 1.) 106 | Idenoised_crop = Idenoised_crop.permute(0, 2, 3, 1)[:, h_min_margin:H-h_max_margin, w_min_margin:W-w_max_margin, :].cpu() 107 | Idenoised_crop = Idenoised_crop.numpy() 108 | 109 | 110 | if opt['save_mat']: 111 | # save denoised data 112 | Idenoised_crop = np.float32(Idenoised_crop) 113 | save_file = os.path.join(opt['mat_dir'], '%04d_%02d.mat' % (i + 1, k + 1)) 114 | sio.savemat(save_file, {'Idenoised_crop': Idenoised_crop}) 115 | 116 | 117 | if __name__ == '__main__': 118 | parser = argparse.ArgumentParser(description="Train the denoiser") 119 | parser.add_argument("--config_file", type=str, default='../option/three_stage.json') 120 | argspar = parser.parse_args() 121 | 122 | opt = parse(argspar.config_file) 123 | opt['mat_dir'] = 'dnd_mat' 124 | opt['save_mat'] = True 125 | opt['ensemble'] = True 126 | recursive_print(opt) 127 | 128 | main(opt) 129 | if opt['save_mat']: 130 | bundle_submissions_srgb(opt['mat_dir']) -------------------------------------------------------------------------------- /submit_test/test_SIDD.py: -------------------------------------------------------------------------------- 1 | import sys 2 | sys.path.append('..') 3 | import argparse 4 | from dataset.SIDD_sRGB import SIDDSrgbBenchmarkDataset 5 | import numpy as np 6 | import os 7 | import scipy.io as sio 8 | from submit_test.ensemble_wrapper import EnsembleWrapper 9 | import torch 10 | from torch.utils.data import DataLoader 11 | from tqdm import tqdm 12 | from utils.option import parse, recursive_print 13 | 14 | def main(opt): 15 | test_set = SIDDSrgbBenchmarkDataset() 16 | test_loader = DataLoader(test_set, batch_size=1) 17 | 18 | if os.path.exists(opt['mat_path']): 19 | os.remove(opt['mat_path']) 20 | 21 | Model = getattr(__import__('model'), opt['model']) 22 | model = Model(opt) 23 | model.data_parallel() 24 | if 'resume_from' in opt: 25 | model.load_model(opt['resume_from']) 26 | if opt['ensemble']: 27 | model = EnsembleWrapper(model) 28 | 29 | count = 0 30 | denoised_block = np.zeros_like(test_set.noisy_block) 31 | for data in tqdm(test_loader): 32 | output = model.validation_step(data) 33 | output = torch.floor(output + 0.5) 34 | output = torch.clamp(output, 0, 255) 35 | output = output.cpu().squeeze(0).permute(1, 2, 0).numpy() 36 | 37 | index_n = count // test_set.noisy_block.shape[1] 38 | index_k = count % test_set.noisy_block.shape[1] 39 | output = np.uint8(output) 40 | denoised_block[index_n, index_k] = output 41 | count += 1 42 | 43 | save_dict = {} 44 | save_dict['__header__'] = b'MATLAB 5.0 MAT-file, Platform: PCWIN64, Created on: Thu Jan 10 13:08:11 2019' 45 | save_dict['__version__'] = 1.0 46 | save_dict['__globals__'] = [] 47 | save_dict['DenoisedBlocksSrgb'] = denoised_block 48 | sio.savemat(opt['mat_path'], save_dict) 49 | 50 | 51 | if __name__ == '__main__': 52 | parser = argparse.ArgumentParser(description="Train the denoiser") 53 | parser.add_argument("--config_file", type=str, default='../option/three_stage.json') 54 | argspar = parser.parse_args() 55 | 56 | opt = parse(argspar.config_file) 57 | opt['mat_path'] = 'SubmitSrgb.mat' 58 | opt['ensemble'] = True 59 | recursive_print(opt) 60 | 61 | main(opt) -------------------------------------------------------------------------------- /utils/build.py: -------------------------------------------------------------------------------- 1 | def build(obj_type, args): 2 | return obj_type(**args) -------------------------------------------------------------------------------- /utils/io.py: -------------------------------------------------------------------------------- 1 | import datetime 2 | import imageio 3 | import numpy as np 4 | import torch 5 | 6 | def date_time(): 7 | now = datetime.datetime.now() 8 | date_time = now.strftime("%Y-%m-%d, %H:%M:%S") 9 | return date_time 10 | 11 | def log(log_file, str, also_print=True, with_time=True): 12 | with open(log_file, 'a+') as F: 13 | if with_time: 14 | F.write(date_time() + ' ') 15 | F.write(str) 16 | if also_print: 17 | if with_time: 18 | print(date_time(), end=' ') 19 | print(str, end='') 20 | 21 | # save numpy image in shape 3xHxW 22 | def np2image(image, image_file): 23 | image = np.transpose(image, (1, 2, 0)) 24 | image = np.clip(image, 0., 1.) 25 | image = image * 255. 26 | image = image.astype(np.uint8) 27 | imageio.imwrite(image_file, image) 28 | 29 | # save tensor image in shape 1x3xHxW 30 | def tensor2image(image, image_file): 31 | image = image.detach().cpu().squeeze(0).numpy() 32 | np2image(image, image_file) 33 | 34 | # return pytorch image in shape 1x3xHxW 35 | def image2tensor(image_file): 36 | image = imageio.imread(image_file).astype(np.float32) / np.float32(255.0) 37 | if len(image.shape) == 3: 38 | image = np.transpose(image, (2, 0, 1)) 39 | elif len(image.shape) == 2: 40 | image = np.expand_dims(image, 0) 41 | image = np.asarray(image, dtype=np.float32) 42 | image = torch.from_numpy(image).unsqueeze(0) 43 | return image 44 | -------------------------------------------------------------------------------- /utils/option.py: -------------------------------------------------------------------------------- 1 | from collections import OrderedDict 2 | import json 3 | from utils.io import log 4 | 5 | def parse(opt_path): 6 | # ---------------------------------------- 7 | # remove comments starting with '//' 8 | # ---------------------------------------- 9 | json_str = '' 10 | with open(opt_path, 'r') as f: 11 | for line in f: 12 | line = line.split('//')[0] + '\n' 13 | json_str += line 14 | 15 | # ---------------------------------------- 16 | # initialize opt 17 | # ---------------------------------------- 18 | opt = json.loads(json_str, object_pairs_hook=OrderedDict) 19 | 20 | return opt 21 | 22 | 23 | def recursive_print(src, dpth=0, key=None): 24 | """ Recursively prints nested elements.""" 25 | tabs = lambda n: ' ' * n * 4 # or 2 or 8 or... 26 | 27 | if isinstance(src, dict): 28 | if key is not None: 29 | print(tabs(dpth) + '%s: ' % (key)) 30 | for key, value in src.items(): 31 | recursive_print(value, dpth + 1, key) 32 | elif isinstance(src, list): 33 | if key is not None: 34 | print(tabs(dpth) + '%s: ' % (key)) 35 | for litem in src: 36 | recursive_print(litem, dpth) 37 | else: 38 | if key is not None: 39 | print(tabs(dpth) + '%s: %s' % (key, src)) 40 | 41 | 42 | def recursive_log(log_file, src, dpth=0, key=None): 43 | """ Recursively prints nested elements.""" 44 | tabs = lambda n: ' ' * n * 4 # or 2 or 8 or... 45 | 46 | if isinstance(src, dict): 47 | if key is not None: 48 | log(log_file, tabs(dpth) + '%s: \n' % (key), with_time=False) 49 | for key, value in src.items(): 50 | recursive_log(log_file, value, dpth + 1, key) 51 | elif isinstance(src, list): 52 | if key is not None: 53 | log(log_file, tabs(dpth) + '%s: \n' % (key), with_time=False) 54 | for litem in src: 55 | recursive_log(log_file, litem, dpth) 56 | else: 57 | if key is not None: 58 | log(log_file, tabs(dpth) + '%s: %s\n' % (key, src), with_time=False) -------------------------------------------------------------------------------- /validate/validate_SIDD.py: -------------------------------------------------------------------------------- 1 | import sys 2 | sys.path.append('..') 3 | import argparse 4 | from skimage.metrics import peak_signal_noise_ratio 5 | import torch 6 | from torch.utils.data import DataLoader 7 | from utils.option import parse, recursive_print 8 | from utils.build import build 9 | 10 | def validate_sidd(model, sidd_loader): 11 | psnrs, count = 0, 0 12 | for data in sidd_loader: 13 | output = model.validation_step(data) 14 | output = torch.floor(output + 0.5) 15 | output = torch.clamp(output, 0, 255) 16 | output = output.cpu().squeeze(0).permute(1, 2, 0).numpy() 17 | gt = data['H'].squeeze(0).permute(1, 2, 0).numpy() 18 | psnr = peak_signal_noise_ratio(output, gt, data_range=255) 19 | psnrs += psnr 20 | count += 1 21 | return psnrs / count 22 | 23 | 24 | def main(opt): 25 | validation_loaders = [] 26 | for validation_dataset_opt in opt['validation_datasets']: 27 | ValidationDataset = getattr(__import__('dataset'), validation_dataset_opt['type']) 28 | validation_set = build(ValidationDataset, validation_dataset_opt['args']) 29 | validation_loader = DataLoader(validation_set, batch_size=1) 30 | validation_loaders.append(validation_loader) 31 | 32 | Model = getattr(__import__('model'), opt['model']) 33 | model = Model(opt) 34 | model.data_parallel() 35 | if 'resume_from' in opt: 36 | model.load_model(opt['resume_from']) 37 | 38 | for validation_loader in validation_loaders: 39 | psnr = validate_sidd(model, validation_loader) 40 | print('%s, psnr: %6.4f' % (validation_loader.dataset.__class__.__name__, psnr)) 41 | 42 | 43 | if __name__ == '__main__': 44 | parser = argparse.ArgumentParser(description="Validate the denoiser") 45 | parser.add_argument("--config_file", type=str, default='../option/three_stage.json') 46 | argspar = parser.parse_args() 47 | 48 | opt = parse(argspar.config_file) 49 | recursive_print(opt) 50 | 51 | main(opt) --------------------------------------------------------------------------------